SYSTEMS AND METHODS FOR AUTOMATIC DATA ANNOTATION

Abstract

Described herein are systems, methods, and instrumentalities associated with automatically annotating a 3D image dataset. The 3D automatic annotation may be accomplished based on a 2D manual annotation provided by an annotator and by propagating, using a set of machine-learning (ML) based techniques, the 2D manual annotation through sequences of 2D images associated with the 3D image dataset. The automatically annotated 3D image dataset may then be used to annotate other 3D image datasets upon passing a readiness assessment conducted using another set of ML based techniques. The automatic annotation of the images may be performed progressively, e.g., by processing a subset or batch of images at a time, and the ML based techniques may be trained to ensure consistency between a forward propagation and a backward propagation.

Description

BACKGROUND

Having annotated data is crucial to the training of machine-learning (ML) models or artificial neural networks. Conventional ways of data annotation rely heavily on manual work (e.g., by qualified annotators such as radiologists if the data includes medical images) and even when computer-based tools are provided, they still require a tremendous amount of human effort (e.g., mouse clicking, drag-and-drop, etc.). This strains resources and often leads to inadequate and/or inaccurate results. Accordingly, it is highly desirable to develop systems and methods to automate the data annotation process such that more data may be obtained for ML training and/or verification.

SUMMARY

Disclosed herein are systems, methods, and instrumentalities associated with automatic 3D data (e.g., 3D images) annotation. According to embodiments of the disclosure, an apparatus configured to perform the data annotation task may include at least one processor that may be configured to obtain a first sequence of two-dimensional (2D) images (e.g., based on a first three-dimensional (3D) image dataset) and further obtain a first manual annotation based on a first user input (e.g., obtained through a graphical user interface provided by the processor), where the first manual annotation may be associate with a first image in the first sequence of 2D images and may indicate a location of a person or an object (e.g., an anatomical structure such as an organ) in the first image. The at least one processor may be configured to annotate, automatically, a first subset of images in the first sequence of 2D images based on the first manual annotation and a first machine-learning (ML) model, and to further annotate, automatically, a second subset of images in the first sequence of 2D images based on the first ML model and a second annotation. The second annotation may be an annotation automatically generated for the last image of the first subset of images or an annotation manually generated for a second image of the first sequence of 2D images. In this manner, the at least one processor may perform the automatic annotation task progressively, e.g., by processing one subset or batch of images at a time.

In some embodiments, the at least one processor may be configured to determine that the number of images included in the first subset of images (e.g., the number of images to be automatically annotated based on the first manual annotation in the first batch) is equal to the size of a pre-defined annotation propagation window. In some embodiments, the at least one processor may be configured to determine that the number of images included in the first subset of images (e.g., the number of images to be automatically annotated based on the first manual annotation in the first batch) is equal to the number of images sequentially located between the first image (e.g., corresponding to the first manual annotation) and the second image (e.g., corresponding to a second manual annotation) in the first sequence of 2D images. In other words, the at least one processor may be configured to automatically annotate images in the first sequence of 2D images based on the first manual annotation and the pre-defined annotation propagation window size until the processor encounters another manually annotated image (e.g., within the propagation window).

In some embodiments, the first ML model may be trained for extracting first features associated with the person or the object from the first manual annotation, extracting respective second features associated with the person or the object from the first subset of images, and automatically annotating the first subset of images based on the first features and the second features. In some embodiments, the at least one processor may be further configured to obtain a third manual annotation that may be associated with a third image in the first subset of images or in the second subset of images, and to annotate, automatically, one or more images adjacent to the third image (e.g., according to the pre-defined annotation propagation window) based on the third manual annotation. This way, a user may adjust an auto-generated annotation and have the adjustment propagated to other images to ensure the quality of the automatic annotation process.

In some embodiments, the first ML model may be trained using a plurality of sequentially ordered training images and, during the training, the first ML model may be used to annotate, automatically, the plurality of sequentially ordered training images in a first order (e.g., an ascending order of image indices) and based on a first manually created training annotation. The first ML model may be further used to annotate, automatically, the plurality of sequentially ordered training images in a second order (e.g., a descending order of the image indices) and based on a second manually created training annotation. The parameters of the first ML model may then be adjusted to enforce consistency between corresponding annotations obtained in the first order and the second order.

In some embodiments, the at least one processor described herein may be further configured to determine, based on a second ML model and a readiness score associated with one or more annotated image sequences, whether to use the one or more annotated image sequences to automatically annotate the second sequence of 2D images. Such a second ML model may be trained for predicting a query annotation based on the one or more annotated image sequences and the readiness score may be determined by comparing the query annotation with a ground truth annotation. If the determination is to use the one or more annotated image sequences to automatically annotate the second sequence of 2D images, the at least one processor may obtain an annotation for the second sequence of 2D images (e.g., an initial annotation that may be propagated through the second sequence of 2D images) based on the one or more annotated image sequences and the second ML model.

BRIEF DESCRIPTION OF THE DRAWINGS

A more detailed understanding of the examples disclosed herein may be had from the following description, given by way of example in conjunction with the accompanying drawing.

FIG. 1 is a diagram illustrating an example of automatic 3D data annotation, in accordance with one or more embodiments of the present disclosure.

FIG. 2 is a diagram illustrating example techniques for automatically annotating a second 3D image dataset based on an annotated first 3D image dataset, in accordance with one or more embodiments of the present disclosure.

FIG. 3 is a diagram illustrating examples machine learning techniques for automatically annotating a second image based on an annotated first image, in accordance with one or more embodiments of the present disclosure.

FIG. 4 is a diagram illustrating examples machine learning techniques for generating an initial annotation for a second 3D image dataset based on an annotated first image dataset, in accordance with one or more embodiments of the present disclosure.

FIG. 5 is a flow diagram illustrating example operations that may be associated with automatically annotating a 3D image dataset, in accordance with one or more embodiments of the present disclosure.

FIG. 6 is a flow diagram illustrating example operations that may be associated with automatically annotating a second 3D image dataset based on an annotated first 3D image dataset, in accordance with one or more embodiments of the present disclosure.

FIG. 7 is a block diagram illustrating an example of progressive annotation in accordance with one or more embodiments of the present disclosure.

FIG. 8 is a block diagram illustrating an example of enforcing consistency between forward automatic annotation and backward automatic annotation in accordance with one or more embodiments of the present disclosure.

FIG. 9 is a flow diagram illustrating example operations that may be associated with training a neural network to perform the tasks described in accordance with one or more embodiments of the present disclosure.

FIG. 10 is a block diagram illustrating example components of an apparatus that may be configured to perform the tasks described in accordance with one or more embodiments of the present disclosure.

DETAILED DESCRIPTION

The present disclosure is illustrated by way of example, and not by way of limitation, in the figures of the accompanying drawings.

FIG. 1 illustrates an example of automatic 3D data annotation in accordance with one or more embodiments of the present disclosure. The example will be described in the context of medical images such as magnetic resonance imaging (MM) images, computed tomography (CT) images, and/or X-ray images, but those skilled in the art will appreciate that the disclosed techniques may also be used to annotate other types of images or data including, for example, red-green-blue (RBG) images, depth images, thermal images, alphanumeric data, etc. As shown in FIG. 1, a first 3D image dataset requiring annotation may include a first sequence of 2D images 102, which may be obtained by splitting the 3D image dataset along a direction (e.g., axis) of the 3D image dataset. For instance, the first 3D image dataset may include MM or CT scans of a first patient captured along one or more axes of an anatomical structure of the patient (e.g., from top to down, from front to back, etc.), and the first sequence of 2D images 102 may be obtained by splitting the first 3D image dataset along one of those axes.

In some embodiments of the present disclosure, a computer-generated user interface (not shown) may be provided to display the first sequence of 2D images 102 (e.g., after they have been obtained using the technique described above), and a user may, through the interface, select and annotate one or more of the first sequence of 2D images 102. For example, the user may select an image 102a from the first sequence of 2D images 102 and annotate, at 104, the image 102a by marking, outlining, or otherwise indicating the location or contour of an object of interest 106 (e.g., a brain hemorrhage) in the image 102a to obtain an annotation 108 (e.g., a segmentation mask) for the original image 102a. The annotation operation at 104 may be performed using tools provided by the user interface, which may allow the user to create the annotation 108 through one or more of a click, a tap, a drag-and-drop, a click-drag-and-release, a sketching or drawing motion, etc. that may be executed by the user with an input device (e.g., a computer mouse, a keyboard, a stylus, a touch screen, the user's finger, etc.). The annotation 108 may then be used at 110 to generate a 3D annotation for the first 3D image dataset, for example, by propagating the annotation 108 through (e.g., by automatically annotating) multiple other images 102b, 102c, etc. of the first sequence 102 (e.g., annotation 108 may be propagated to all or a subset of the first sequence of 2D images 102) to obtain annotations (e.g., segmentation masks) 108b, 108c, etc. The automatic 3D annotation at 110 may be accomplished using a first machine-learning (ML) data annotation model, which may be trained for detecting features associated with the object of interest 106 in image 102a (e.g., based on manual annotation provided at 104), identifying areas having similar features from the other 2D images (e.g., 102b, 102c, etc.) of the first sequence 102, and automatically annotating those areas as containing the object of interest 106. The implementation and/or training of the first ML data annotation model will be described in greater detail below, and the term “machine-learning model” may be used interchangeably herein with the term “machine-learned model” or “artificial intelligence model.”

The 3D annotation generated at 110 for the first 3D image dataset may be displayed to a user (e.g., the same user who created the 2D annotation 108), who may then confirm and/or adjust the 3D annotation (e.g., through one or more user inputs), for example, using the same user interface described above. The confirmed and/or adjusted 3D annotation may be used to automatically annotate (e.g., generate a 3D annotation for) other 3D image datasets such as a second 3D image dataset that may be associated with a second patient, as described below. The user-provided adjustment may also be used to improve the first ML data annotation model, for example, through reinforcement learning, which may be conducted (e.g., in an online manner) after the first ML data annotation model has been trained (e.g., in an offline manner) and deployed. The improvement may be accomplished, for example, based on the differences (e.g., prediction errors) between the automatic annotation predicted by the first ML data annotation model and the user input.

FIG. 2 illustrates an example of automatically annotating a second 3D image dataset based on an annotated first 3D image dataset such as the 3D image dataset 102 of FIG. 1. Similar to the first 3D image dataset 102, the second 3D image dataset may include a second sequence of 2D images 202, which may be obtained by splitting the second 3D image dataset along a direction (e.g., axis) of the second 3D image dataset. For instance, the 3D image dataset may include MRI or CT scans of the second patient captured along one or more axes of an anatomical structure (e.g., the same anatomical structure in the first 3D image dataset) of the patient (e.g., from top to down, from front to back, etc.), and the second sequence of 2D images 202 may be obtained by splitting the second 3D image dataset along one of those axes (e.g., along the same axis from which the first 3D image dataset is obtained).

To annotate the second 3D image dataset based on the first 3D image dataset such as a previously annotated sequence of 2D images (e.g., the first sequence 102 of FIG. 1), one or more images 202a of the second sequence of 2D images 202 may be automatically selected, for example, based on similarities between the image(s) 202a and one or more images of the previously annotated sequence (e.g., an image 202a from the second sequence 202 may be selected for being similar to one of the images in the previously annotated sequence). Once the similar image(s) from the second sequence 202 are selected, initial annotation(s) (e.g., 2D annotation(s)) may be automatically generated (e.g., as part of the operation at 206) based on the similarity between the selected image(s) from sequence 202 and the corresponding image(s) in the previously annotated sequence, and the annotation(s) (e.g., segmentation mask 204a, etc.) that have been generated for the corresponding image(s). One or more of these tasks may be accomplished using a second ML data annotation model, which may be trained for identifying similar images from the second sequence 202 and the annotated sequence 204, and generating the initial annotation(s) for the second sequence 202, as described above. The implementation and training of the second ML data annotation model will be described in greater detail below.

The automatically generated initial 2D annotation(s) for the second sequence 202 may be confirmed and/or adjusted by a user, for example, using the user interface described herein. The confirmed or adjusted 2D annotation may then be propagated (e.g., from image 202a) to other images (e.g., 202b, 202c, etc.) of the second sequence 202 (e.g., to all or a subset of the second sequence of 2D images 202) to obtain a 3D annotation 208 for the second 3D image dataset (e.g., comprising segmentation masks 208a, 208b, 208c, etc.). The propagation (e.g., the automatic annotation of images 202b, 202c, etc.) may be accomplished, for example, based on the first ML data annotation model described herein. The user-confirmed or adjusted annotation(s) may also be used for improving the second ML data annotation model, for example, to generate more accurate initial annotation(s) for subsequent 3D image datasets.

FIG. 3 illustrates an example of an ML model (e.g., the first ML data annotation described herein) for automatically annotating a second image 304 of an object based on a first image 302 of the object and an annotation (e.g., a segmentation mask) 306 already generated for the first image 302. As shown, the ML model may include one or more features extraction modules 308 configured to extract features, f₁, f₂, v₁ from the first image 302, the second image 304, and the annotation (e.g., segmentation mask) 306, respectively. The feature extraction module 308 may be implemented using an artificial neural network such as a convolutional neural network (CNN). In examples, such a CNN may include an input layer configured to receive an input image (e.g., including a segmentation mask) and one or more convolutional layers, pooling layers, and/or fully-connected layers configured to process the input image. Each of the convolutional layers may include a plurality of convolution kernels or filters with respective weights, the values of which may be learned through a training process such that features associated with an object of interest in the image may be identified using the convolution kernels or filters upon completion of the training. The convolutional layers may be followed by batch normalization and/or linear or non-linear activation (e.g., such as a rectified linear unit (ReLU) activation function), and the features extracted through the convolution operations may be down-sampled through one or more pooling layers to obtain a representation of the features, for example, in the form of a feature vector or a feature map (e.g., feature representation f₁, f₂, or v₁).

In some implementations of the ML model shown in FIG. 3, the same feature extraction module may be used to encode the features from the first image 302, the second image 304, and the segmentation mask 306. In other implementations, different feature extraction modules may be used to encode the features from the images. For example, a key encoder may be used to encode features f₁ and f₂ from the first image 302 and the second image 304, respectively, and a value encoder may be used to encode features v₁ from the segmentation mask 306. In examples, image features f₁ and f₂ may be used to derive (e.g., calculate or learn) an affinity matrix 310 that may represent the spatial-temporal correspondence between the first image 302 and the second image 304. The affinity matrix 310 and the features v₁ of the segmentation mask 306 may then be used to decode (e.g., at 312) a segmentation mask 314 for the second image 304. The feature decoding at 312 may be conducted using a CNN that may include one or more un-pooling layers and one or more transposed convolutional layers. Through the un-pooling layers, the features provided by the affinity matrix 310 and/or features v₁ of the segmentation mask 306 may be up-sampled, and the up-sampled features may be further processed through the one or more transposed convolutional layers (e.g., via a plurality of deconvolution operations) to derive an up-scaled or dense feature map or feature vector. The dense feature map or vector may then be used to predict areas (e.g., pixels) in the second image 304 that may belong to object of interest. The prediction may be represented by the segmentation mask 314, which may include a respective probability value (e.g., ranging from 0 to 1) for each image pixel indicating whether the image pixel belongs to the object of interest (e.g., having a probability value above a preconfigured threshold) or a background area (e.g., having a probability value below a preconfigured threshold).

FIG. 4 illustrates an example of an ML model (e.g., the second ML data annotation model described herein) for automatically annotating a second image 404 (e.g., an image from sequence 202 of FIG. 2) of an object based on a first image 402 of the object and an annotation 406 for the first image 402. The first image 402 may be an image from a previously annotated sequence (e.g., image 102a of FIG. 1) and the annotation 406 (e.g., annotation 108a of FIG. 1) may be manually generated (e.g., with human interventions), for example, with the user interface and/or tools described herein. Based on the first image 402 and the annotation 406 (e.g., which may be a segmentation mask), a first plurality of features, f₁, associated with the annotated object of interest may be extracted from the first image 402 and/or the annotation 406 using an extraction module 408. The extraction module may be learned and/or implemented using an artificial neural network such as a convolutional neural network (CNN). In examples, such a CNN may include an input layer configured to receive an input image and one or more convolutional layers, pooling layers, and/or fully-connected layers configured to process the input image. Each of the convolutional layers may include a plurality of convolution kernels or filters with respective weights, the values of which may be learned through a training process such that features associated with an object of interest in the image may be identified using the convolution kernels or filters upon completion of the training. The convolutional layers may be followed by batch normalization and/or linear or non-linear activation (e.g., such as a rectified linear unit or ReLU activation function), and the features extracted through the convolution operations may be down-sampled through one or more pooling layers to obtain a representation of the features, for example, in the form of a feature vector or a feature map. In some examples (e.g., if a segmentation mask for the input image is to be generated), the CNN may also include one or more un-pooling layers and one or more transposed convolutional layers. Through the un-pooling layers, the network may up-sample the features extracted from the input image and process the up-sampled features through the one or more transposed convolutional layers (e.g., via a plurality of deconvolution operations) to derive an up-scaled or dense feature map or feature vector. The dense feature map or vector may then be used to predict areas (e.g., pixels) in the input image that may belong to object of interest. The prediction may be represented by a mask, which may include a respective probability value (e.g., ranging from 0 to 1) for each image pixel that indicates whether the image pixel may belong to object of interest (e.g., having a probability value above a preconfigured threshold) or a background area (e.g., having a probability value below a preconfigured threshold).

The annotation 406 for the first image 402 may be used to enhance the completeness and/or accuracy of the first plurality of features f₁ (e.g., which may be obtained as a feature vector or feature map). For example, using a normalized version of the annotation 406 (e.g., by converting probability values in the annotation mask to a value range between 0 and 1), the first image 402 (e.g., pixel values of the first image 402) may be weighted (e.g., before the weighted imagery data is passed to the feature extraction operation at 408) such that pixels belonging to the object of interest may be given larger weights during the feature extraction process. As another example, the normalized annotation mask may be used to apply respective weights to the features (e.g., preliminary features) extracted at 408 such that features associated with the object of interest may be given larger weights within the feature representation f₁.

The second image 404, which may include the same object of interest as the first image 402, may be processed through a feature extraction module 410 (e.g., which may be the same feature extraction module as 408 or a difference feature extraction module) to determine a second plurality of features f₂. The second plurality of features f₂ may be represented in the same format as the first plurality of features f₁, (e.g., as a feature vector) and/or may have the same size as f₁. The two sets of features may be used jointly to determine a set of informative features f₃ that may be indicative of the pixel characteristics of the object of interest in first image 402 and the second image 404. For instance, informative features f₃ may be obtained by comparing features f₁ and f₂, and selecting the features that are common to both f₁ and f₂. One example way of accomplishing this task may be to normalize the feature vectors of f₁ and f₂ (e.g., such that both vectors have values ranging from 0 to 1), compare the two normalized vectors (e.g., based on (f₁−f₂)), and select corresponding elements in the two vectors that have a value difference smaller than a predefined threshold as the informative features f₃.

In examples, the second plurality of features f₂ extracted from the second image 404 and/or the informative features f₃ may be further processed at 412 to gather information (e.g., from certain dimensions of f₂) that may be used to automatically annotate the object of interest in the second image 404. For example, based on the informative features f₃, an indicator vector having the same size as the feature vector f₁ and/or f₂ may be derived in which elements that correspond to informative features f₃ may be given a value of 1 and the remaining elements may be given a value of 0. A score may then be calculated to aggregate of the informative features f₃ and/or the informative elements of feature vector f₂. Such a score may be calculated, for example, by conducting an element-wise multiplication of the indicator vector and feature vector f₂. Using the calculated score, an annotation 414 of the object of interest may be automatically generated for the second image 404, for example, by backpropagating a gradient of the score through the first ML data annotation model (e.g., through the neural network used to implement the first ML data annotation model) and determining pixel locations (e.g., spatial dimensions) that may correspond to the object of interest based on the gradient values associated with the pixel locations. For instance, pixel locations having positive gradient values during the backpropagation (e.g., these pixel locations may make positive contributions to the desired result) may be determined to be associated with the object of interest and pixel locations having negative gradient values during the backpropagation (e.g., these pixel locations may not make contributions or may make negative contributions to the desired result) may be determined to be not associated with the object of interest. Annotation 414 for the second image 404 may then be generated based on a weighted linear combination of the feature maps or feature vectors obtained by the first ML data annotation model (e.g., the gradients may operate as the weights in the linear combination).

The annotation (e.g., annotation 414) automatically generated using the techniques described above may be presented to a user, for example, through the user interface described herein so that adjustments may be made by the user to refine the annotation. For example, the user interface may allow the user to adjust the annotation 414 by executing one or more of a click, a tap, a drag-and-drop, a click-drag-and-release, a sketching or drawing motion, etc. Adjustable control points may be provided along the contour of the annotation 414 and the user may be able to change the shape of the annotation 414 by manipulating one or more of these control points (e.g., by dragging and dropping the control points to various new locations on the display screen).

FIG. 5 illustrates example operations 500 that may be associated with the automatic annotation of a 3D image dataset. As shown, the operations 500 may include obtaining, at 502, a first sequence of 2D images associated with a first 3D image dataset. The first sequence of 2D images may be obtained, for example, by splitting the 3D image dataset along an axis of the 3D image dataset and including the 2D images associated with the axis in the first sequence. The operations 500 may further include receiving, at 504, an annotation of a first 2D image in the first sequence of 2D images, where the annotation may be created by an annotator using the user interface and/or annotation tools described herein. Based on the received annotation of the first 2D image, a 3D annotation may be generated for the first 3D image dataset based on a ML data annotation model, for example, by propagating the annotation of the first 2D image to multiple other 2D images of the first sequence of 2D images. The annotation of the first 2D image may indicate (e.g., delineate or segment) an object of interest in the first 2D image, and the first ML data annotation model may be trained for detecting features associated with the object of interest in the other 2D images of the first sequence and automatically annotating the other 2D images based on the detected features.

The automatically generated 3D annotation for the first 3D image dataset may be confirmed and/or adjusted by a user, and the parameters of the first ML model may be adjusted based on the user adjustment, for example, through reinforcement learning (e.g., conducted in an online manner). The confirmed or adjusted 3D annotation may be used to automatically annotate other 3D image datasets including, e.g., a second 3D image dataset associated with a second patient.

FIG. 6 illustrates example operations 600 that may be associated with the automatic annotation of a second 3D image dataset based on an annotated first 3D image dataset. As shown, the operations 600 may include obtaining, at 602, the second 3D image dataset and splitting the second 3D image dataset (e.g., along an axis of the second 3D image dataset) to obtain a second sequence of 2D images at 604. From the second sequence of 2D images, a 2D image may be selected at 606 based on a similarity between the 2D image and a corresponding 2D image in the annotated first 3D image dataset. An initial 2D annotation may then be automatically generated for the 2D image in the second sequence at 608 based on the annotation of the corresponding 2D image in the first 3D image dataset. The initial 2D annotation may be confirmed or adjusted (e.g., by a human annotator) at 610 before the confirmed or adjusted 2D annotation is used to generate a 3D annotation for the second 3D image dataset at 612, e.g., by propagating the confirmed or adjusted 2D annotation through multiple other 2D images of the second sequence of 2D images based on the first ML data annotation model described herein.

The automatic annotation operations described herein may be performed in a progressive manner, for example, to reduce the amount of human efforts and/or computational resources (e.g., memory consumptions) involved. FIG. 7 illustrates an example of a progressive annotation process during which images from a 3D image dataset may be automatically annotated in batches and/or based on interactions with a human annotator. As shown in the example of FIG. 7, a 3D image dataset (e.g., obtained from an imaging sensor as described herein) may be split into sequences of 2D images along respective axes (e.g., orthogonal axes) and one or more manual annotations (e.g., user-created annotations) may be obtained for the sequences of 2D images based on user inputs received via the user interface or the annotation tools described herein. The manual annotations may include, for example, manual annotation 1 and manual annotation 2 shown in FIG. 7, which may be respectively created for images 1 and m of a first sequence of 2D images (e.g., sequentially ordered or index as between 1 and n). Rather than annotating the sequence of 2D images all at once based on manual annotation 1 and/or manual annotation 2, the example approach illustrated by FIG. 7 may divide the annotation task into k batches or sub-parts, and process one batch or sub-part at a time to reduce the burden on human and computational resources. With each batch or sub-part, a subset of images from the first sequence of 2D images may be loaded into a memory pool and automatically annotated based on a manual annotation or an automatic annotation generated from a previous batch or sub-part. The number of images included in each batch or sub-part may be determined based on an annotation propagation window size (e.g., with a pre-defined and adjustable value) or the availability of a manual annotation within the annotation propagation window. For example, with a pre-defined annotation propagation window size of w, manual annotation 1 (e.g., a middle one of the available manual annotations) associated with image 1 may be used as a basis to annotate w images that may be adjacent to image 1 (e.g., in front of or subsequent to image 1) in a first batch, e.g., using the ML data annotation model described herein. Then, using the (w+1)-th image in front of or subsequent to image 1 as a basis, a second batch of images may be annotated automatically in a similar manner. These operations may be repeated till the end(s) of the first sequence of 2D images is reached and the 2D annotations thus obtained may be used to generate a 3D annotation (e.g., an initial 3D annotation) for the 3D image database.

In examples, if there is a manual annotation within a current annotation propagation window, the corresponding batch may end before the manual annotation and the manual annotation may be set/used as a reference for the next batch. For example, as shown in FIG. 7, batch 2 may process images leading up to the image (e.g., image m) associated with manual annotation 2, even if the number of images processed in the batch is less than annotation propagation window size w, and manual annotation 2 may be used as the basis for processing the next batch (e.g., batch 3). In this way, the basis or reference for a new automatic batch may be reset or re-aligned with (e.g., periodically) a user-provided annotation (e.g., which may have a higher level of accuracy), thus preventing potential errors that may have occurred during a previous batch from spreading to the new batch.

In examples, a user (e.g., a human annotator) may adjust an auto-generated annotation for image i using the interface or tools described herein, and the adjusted annotation may be used a reference or basis for annotating (e.g., re-annotating) a batch of other images located adjacent to image i, where the number of the images included in the batch may be equal to the smaller of annotation propagation window size w described herein or the number of images located between image i and a manually annotated image nearest to image i. The last annotation in this batch (or the nearest manual annotation) may then be used as a basis for annotating another batch of images and the process may be repeated till a satisfactory 3D annotation is derived for the 3D dataset.

It should be noted that image 1 shown in FIG. 7 may not necessarily be the first image in the first sequence of 2D images. Rather, the index “1” may merely indicate that a manual annotation generated based on image 1 may be used as a basis for automatically annotating other images (e.g., within a batch). It should also be noted that the automatic annotation process described herein may be performed in either a forward direction (e.g., based on an ascending order of the image indices starting from a base image) and/or a backward direction (e.g., based on a descending order of the image indices starting from a base image). The training of the ML data annotation model described herein may, in examples, be conducted to enforce consistency between automatic annotations performed in the forward direction and those performed in the backward direction. FIG. 8 illustrates such an example.

As shown in FIG. 8, the ML data annotation model described herein may be trained using a sequence of 2D images (e.g., images 1 through n) and one or more manual annotations (e.g., manual annotation 1 and/or 2). During the training, the ML data annotation model may be used to annotate, automatically, the sequence of 2D images in a first direction (e.g., a forward direction from image 1 to image n) based on a first manual annotation (e.g., manual annotation 1 associated with image 1) and to annotate, automatically, the sequence of 2D images in a second direction (e.g., a backward direction from image n to image 1) based on a second manual annotation (e.g., manual annotation 2 associated with image n). A cycle consistency constraint (e.g., in addition to the ground truth-based constraint(s) described herein) may be imposed on the ML data annotation model such that the annotations obtained in the first direction (e.g., starting from image 1 and ending at image n) are consistent with the annotations obtained in the second direction (e.g., starting from image n and ending at image 1). In examples, the consistency constraint may be performed on a subset of images (e.g., any number of images such as three or more images) that may be randomly selected from the 2D image sequence. In these manners, once the ML data annotation model is trained, it may be used to propagate any user-annotated image within an image sequence (e.g., at the beginning of the sequence, the end of the sequence, in the middle of the sequence, etc.) to the rest of the sequence and may do so in a forward or a backward direction.

Annotations obtained for a first 3D image set or a first set of one or more 2D image sequences (e.g., from the first 3D image dataset) may be used (e.g., as reference or support annotations) to automatically annotate a second 3D image set or a second sets of one or more 2D image sequence. In examples, the first 3D image set or the first set of one or more 2D image sequences may belong to a first patient, the second 3D image set or the second set of 2D image sequences may belong to a second patient, and the automatic annotation may be referred to herein as cross-sequence annotation. Such a cross-sequence annotation task may be performed based on an ML model (e.g., referred to herein as a cross-sequence annotation ML model) and a readiness check may be performed (e.g., automatically) on the first 3D image set or the first set of one or more 2D image sequences before it is used to annotate the second 3D image set or the second set of 2D image sequences (e.g., to derive an initial annotation that may be propagated through the second 3D image set or the second set of 2D image sequences).

The cross-sequence annotation ML model may be implemented using a branch of the ML data annotation neural network described herein or using a separate neural network. In examples, such a cross-sequence annotation ML model may employ a two-branch architecture. A first branch of the ML model may be configured to receive one or more annotated image sets or image sequences (e.g., referred to herein as support annotations) as an input, while a second branch of the ML model may be configured to receive an un-annotated image set or image sequence as an input. From these inputs, the cross-sequence annotation ML model may predict an annotation for the un-annotated image set or image sequence (e.g., an initial annotation that may be propagated through the un-annotated image set or image sequences) based on features extracted from the support annotations. In doing so, the cross-sequence annotation ML model may assess the readiness of the support annotations (e.g., by calculating a readiness score for the support annotations) and proceed with the automatic, cross-sequence annotation if (e.g., only if) the readiness of the support annotations exceeds a pre-defined threshold (e.g., the value of this threshold may be determined based on various factors such as the size of the region to be annotated). If the readiness of the support annotations is lower than the pre-defined threshold, a manual annotation may be obtained (e.g., from a human annotator) and used as a basis for annotating the un-annotated image set or image sequence.

The training of the cross-sequence annotation ML model (e.g., a neural network implementing the cross-sequence annotation ML model) may be conducted based on a training dataset comprising N annotated (e.g., manually annotated) image sequences (e.g., the value of N may vary based on factors such as the quality of the training images). During the training, the cross-sequence annotation ML model may be used to predict, for each image sequence i of the N image sequences, an annotation (e.g., referred to herein as a query annotation) for the image sequence (e.g., an annotation for one of the images in the image sequence) based on the other (N−1) annotated image sequences. The cross-sequence annotation ML model may predict the query annotation, for example, by extracting features from the other (N−1) annotated image sequences and inferring the query annotation for image sequence i based on those extracted features (e.g., based on an average or a maximums of the respective sets of features extracted from the (N−1) annotated image sequences). The cross-sequence annotation ML model may then compare the predicted query annotation for image sequence i with a corresponding manual annotation already available in the training dataset (e.g., a ground truth annotation) and calculate a readiness score based on the comparison.

The readiness score may be calculated in different ways including, but not limited to, by calculating an intersection-over-union (IoU) between the predicted query annotation and the ground truth annotation, by determining a ratio of true positive points in the predicted query annotation to a sum of true positive points and false positive points in the predicted annotation, based on the correctness metric associated with the predicted query annotation, etc. The same operations may be applied to each of the N image sequences, through which N readiness scores may be obtained and an annotation may be generated (e.g., based on an average or maximum of the annotations predicted for each image sequence i) for annotating another 3D image set or 2D image sequence. An overall readiness score representing the readiness of the N image sequences for the cross-sequence annotation task may then be calculated, for example, based on an average or median (or other statistical summary) of the N readiness scores. If the overall readiness score is satisfactory (e.g., based on an empirical evaluation), the N image sequences may be determined to be ready for the cross-sequence annotation task (e.g., for generating an initial annotation for a new image sequence). Otherwise (e.g., if the overall readiness score is unsatisfactory), more annotated image sequences (e.g., a (N+1)-th annotated image sequence) may be added to enrich the support annotation pool (e.g., comprising the annotated image sequences) before the pool of images can be used for the cross-sequence annotation task.

The images (e.g., the N image sequences) used to train the cross-sequence annotation ML model may be pre-processed to increase the diversity of the training data and the robustness of the ML model. The pre-processing may be performed, for example, by applying various mathematical operations to the training images including, but not limited to, cropping, rotation, affine transformation, etc. The annotation generated by the cross-sequence annotation ML model may be in different formats, such as, e.g., a binary mask (e.g., in which 0 may represent background and 1 may represent the object of interest), a bounding box (e.g., surrounding the object of the interest in an image), or a group of positive and negative seeds/points that may be used for mask inference.

FIG. 9 illustrates example operations 900 that may be associated with training a neural network (e.g., a neural network used to implement the ML data annotation models described herein) to perform one or more of the tasks described herein. As shown, the training operations may include initializing the operating parameters of the neural network (e.g., weights associated with various layers of the neural network) at 902, for example, by sampling from a probability distribution or by copying the parameters of another neural network having a similar structure. The training operations may further include processing an input (e.g., a training image) using presently assigned parameters of the neural network at 904, and making a prediction for a desired result (e.g., an estimated annotation) at 906. The prediction result may be compared to a ground truth at 908 to determine a loss associated with the prediction, for example, based on a loss function such as mean squared errors between the prediction result and the ground truth, an L1 norm, an L2 norm, etc. At 910, the loss may be used to determine whether one or more training termination criteria are satisfied. For example, the training termination criteria may be determined to be satisfied if the loss is below a threshold value or if the change in the loss between two training iterations falls below a threshold value. If the determination at 910 is that the termination criteria are satisfied, the training may end; otherwise, the presently assigned network parameters may be adjusted at 912, for example, by backpropagating a gradient descent of the loss function through the network before the training returns to 906.

For simplicity of explanation, the training operations are depicted and described herein with a specific order. It should be appreciated, however, that the training operations may occur in various orders, concurrently, and/or with other operations not presented or described herein. Furthermore, it should be noted that not all operations that may be included in the training method are depicted and described herein, and not all illustrated operations are required to be performed.

The systems, methods, and/or instrumentalities described herein may be implemented using one or more processors, one or more storage devices, and/or other suitable accessory devices such as display devices, communication devices, input/output devices, etc. FIG. 10 is a block diagram illustrating an example apparatus 1000 that may be configured to perform the automatic image annotation tasks described herein. As shown, apparatus 1000 may include a processor (e.g., one or more processors) 1002, which may be a central processing unit (CPU), a graphics processing unit (GPU), a microcontroller, a reduced instruction set computer (RISC) processor, application specific integrated circuits (ASICs), an application-specific instruction-set processor (ASIP), a physics processing unit (PPU), a digital signal processor (DSP), a field programmable gate array (FPGA), or any other circuit or processor capable of executing the functions described herein. Apparatus 1000 may further include a communication circuit 1004, a memory 1006, a mass storage device 1008, an input device 1010, and/or a communication link 1012 (e.g., a communication bus) over which the one or more components shown in the figure may exchange information.

Communication circuit 1004 may be configured to transmit and receive information utilizing one or more communication protocols (e.g., TCP/IP) and one or more communication networks including a local area network (LAN), a wide area network (WAN), the Internet, a wireless data network (e.g., a Wi-Fi, 3G, 4G/LTE, or 5G network). Memory 1006 may include a storage medium (e.g., a non-transitory storage medium) configured to store machine-readable instructions that, when executed, cause processor 1002 to perform one or more of the functions described herein. Examples of the machine-readable medium may include volatile or non-volatile memory including but not limited to semiconductor memory (e.g., electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM)), flash memory, and/or the like. Mass storage device 1008 may include one or more magnetic disks such as one or more internal hard disks, one or more removable disks, one or more magneto-optical disks, one or more CD-ROM or DVD-ROM disks, etc., on which instructions and/or data may be stored to facilitate the operation of processor 1002. Input device 1010 may include a keyboard, a mouse, a voice-controlled input device, a touch sensitive input device (e.g., a touch screen), and/or the like for receiving user inputs to apparatus 1000.

It should be noted that apparatus 1000 may operate as a standalone device or may be connected (e.g., networked, or clustered) with other computation devices to perform the functions described herein. And even though only one instance of each component is shown in FIG. 10, a skilled person in the art will understand that apparatus 1000 may include multiple instances of one or more of the components shown in the figure.

While this disclosure has been described in terms of certain embodiments and generally associated methods, alterations and permutations of the embodiments and methods will be apparent to those skilled in the art. Accordingly, the above description of example embodiments does not constrain this disclosure. Other changes, substitutions, and alterations are also possible without departing from the spirit and scope of this disclosure. In addition, unless specifically stated otherwise, discussions utilizing terms such as “analyzing,” “determining,” “enabling,” “identifying,” “modifying” or the like, refer to the actions and processes of a computer system, or similar electronic computing device, that manipulates and transforms data represented as physical (e.g., electronic) quantities within the computer system's registers and memories into other data represented as physical quantities within the computer system memories or other such information storage, transmission or display devices.

It is to be understood that the above description is intended to be illustrative, and not restrictive. Many other implementations will be apparent to those of skill in the art upon reading and understanding the above description. The scope of the disclosure should, therefore, be determined with reference to the appended claims, along with the full scope of equivalents to which such claims are entitled.

Claims

1. An apparatus, comprising: at least one processor configured to: obtain a first sequence of two-dimensional (2D) images;obtain a first manual annotation based on a first user input, wherein the first manual annotation is associate with a first image of the first sequence of 2D images and indicates a location of a person or an object in the first image;annotate, automatically, a first subset of images in the first sequence of 2D images based on the first manual annotation and a first machine-learning (ML) model; andannotate, automatically, a second subset of images in the first sequence of 2D images based on the first ML model and a second annotation associated with a second image of the first sequence of 2D images, wherein the second annotation is automatically generated based on the first manual annotation or manually generated based on a second user input, the second annotation indicating the location of the person or the object in the second image.

2. The apparatus of claim 1, wherein the at least one processor being configured to automatically annotate the first subset of images based on the first manual annotation comprises the at least one processor being configured to determine that a number of images to be automatically annotated based on the first manual annotation is equal to a pre-defined annotation propagation window size, and wherein the second annotation corresponds to an annotation automatically generated for a last image of first subset of images.

3. The apparatus of claim 1, wherein the second annotation corresponds to an annotation manually generated for the second image, and wherein the at least one processor being configured to automatically annotate the first subset of images based on the first manual annotation comprises the at least one processor being configured to determine that a number of images to be automatically annotated based on the first manual annotation is equal to a number of images sequentially located between the first image and the second image.

4. The apparatus of claim 1, wherein the first ML model is trained for extracting first features associated with the person or the object from the first manual annotation, extracting respective second features associated with the person or the object from the first subset of images, and automatically annotating the first subset of images based on the first features and the second features.

5. The apparatus of claim 1, wherein the at least one processor is further configured to obtain a third manual annotation that is associated with a third image in the first subset of images or in the second subset of images and to annotate, automatically, one or more images adjacent to the third image based on the third manual annotation, the third manual annotation indicating the location of the person or the object in the third image.

6. The apparatus of claim 1, wherein the first ML model is trained using a plurality of sequentially ordered training images and wherein, during the training of the first ML model: the first ML model is used to annotate, automatically, the plurality of sequentially ordered training images in a first order and based on a first training annotation;the first ML model is further used to annotate, automatically, the plurality of sequentially ordered training images in a second order and based on a second training annotation; andparameters of the first ML model are adjusted to reduce a difference between annotations obtained in the first order and corresponding annotations obtained in the second order.

7. The apparatus of claim 6, wherein the first order is based on an ascending order of image indices associated with the plurality of sequentially ordered training images and the second order is based on a descending order of the image indices associated with the plurality of sequentially ordered training images.

8. The apparatus of claim 1, wherein the at least one processor is further configured to: obtain a second sequence of 2D images;determine, based on a second ML model and a readiness score associated with one or more annotated image sequences, whether to use the one or more annotated image sequences to automatically annotate the second sequence of 2D images, wherein the second ML model is trained for predicting a query annotation based on the one or more annotated image sequences and wherein the readiness score is determined by comparing the query annotation with a ground truth annotation; andbased on a determination to use the one or more annotated image sequences to automatically annotate the second sequence of 2D images, obtain an annotation for the second sequence of 2D images based on the one or more annotated image sequences and the second ML model.

9. The apparatus of claim 8, wherein the one or more annotated image sequences are associated with a first patient and wherein the second sequence of 2D images is associated with a second patient.

10. The apparatus of claim 1, wherein the at least one processor is further configured to provide a graphical user interface for obtaining the first user input or the second user input.

11. A method of automatic image annotation, the method comprising: obtaining a first sequence of two-dimensional (2D) images;obtaining a first manual annotation based on a first user input, wherein the first manual annotation is associate with a first image of the first sequence of 2D images and indicates a location of a person or an object in the first image;annotating, automatically, a first subset of images in the first sequence of 2D images based on the first manual annotation and a first machine-learning (ML) model; andannotating, automatically, a second subset of images in the first sequence of 2D images based on the first ML model and a second annotation associated with a second image of the first sequence of 2D images, wherein the second annotation is automatically generated based on the first manual annotation or manually generated based on a second user input, the second annotation indicating the location of the person or the object in the second image.

12. The method of claim 11, wherein annotating, automatically, the first subset of images based on the first manual annotation comprises determining that a number of images to be automatically annotated based on the first manual annotation is equal to a pre-defined annotation propagation window size, and wherein the second annotation corresponds to an annotation automatically generated for a last image of first subset of images.

13. The method of claim 11, wherein the second annotation corresponds to an annotation manually generated for the second image, and wherein annotating, automatically, the first subset of images based on the first manual annotation comprises determining that a number of images to be automatically annotated based on the first manual annotation is equal to a number of images sequentially located between the first image and the second image.

14. The method of claim 11, wherein the first ML model is trained for extracting first features associated with the person or the object from the first manual annotation, extracting respective second features associated with the person or the object from the first subset of images, and automatically annotating the first subset of images based on the first features and the second features.

15. The method of claim 11, further comprising obtaining a third manual annotation that is associated with a third image in the first subset of images or in the second subset of images and annotating, automatically, one or more images adjacent to the third image based on the third manual annotation, wherein the third manual annotation indicates the location of the person or the object in the third image.

16. The method of claim 11, wherein the first ML model is trained using a plurality of sequentially ordered training images and wherein, during the training of the first ML model: the first ML model is used to annotate, automatically, the plurality of sequentially ordered training images in a first order and based on a first training annotation;the first ML model is further used to annotate, automatically, the plurality of sequentially ordered training images in a second order and based on a second training annotation; andparameters of the first ML model are adjusted to reduce a difference between annotations obtained in the first order and corresponding annotations obtained in the second order.

17. The method of claim 16, wherein the first order is based on an ascending order of image indices associated with the plurality of sequentially ordered training images and the second order is based on a descending order of the image indices associated with the plurality of sequentially ordered training images.

18. The method of claim 11, further comprising: obtaining a second sequence of 2D images;determining, based on a second ML model and a readiness score associated with one or more annotated image sequences, whether to use the one or more annotated image sequences to automatically annotate the second sequence of 2D images, wherein the second ML model is trained for predicting a query annotation based on the one or more annotated image sequences and wherein the readiness score is determined by comparing the query annotation with a ground truth annotation; andbased on a determination to use the one or more annotated image sequences to automatically annotate the second sequence of 2D images, obtaining an annotation for the second sequence of 2D images based on the one or more annotated image sequences and the second ML model.

19. The method of claim 18, wherein the one or more annotated image sequences are associated with a first patient and wherein the second sequence of 2D images is associated with a second patient.

20. A non-transitory computer-readable medium comprising instructions that, when executed by a processor included in a computing device, cause the processor to: obtain a first sequence of two-dimensional (2D) images;obtain a first manual annotation based on a first user input, wherein the first manual annotation is associate with a first image of the first sequence of 2D images and indicates a location of a person or an object in the first image;annotate, automatically, a first subset of images in the first sequence of 2D images based on the first manual annotation and a first machine-learning (ML) model; andannotate, automatically, a second subset of images in the first sequence of 2D images based on the first ML model and a second annotation associated with a second image of the first sequence of 2D images, wherein the second annotation is automatically generated based on the first manual annotation or manually generated based on a second user input, the second annotation indicating the location of the person or the object in the second image.