Identifying anatomical features

Abstract

In one embodiment, a computer-implemented method 100 is described. The method 100 includes receiving 102 imaging data representative of a volume of an anatomical structure of a subject. The received imaging data includes at least one unidentified anatomical feature of interest. The method 100 further includes identifying 104 the anatomical feature of interest in the received imaging data using a machine learning model configured to perform a segmentation technique to identify the at least one anatomical feature of interest in the received imaging data.

Description

[Technical field]

[0001] The present invention relates to a method, a non-transitory machine-readable medium and an apparatus for identifying anatomical features. [Background technology]

[0002] Magnetic resonance imaging (MRI) and other radiological imaging techniques for imaging the heart can be used to provide data to facilitate quantification of ventricular volumes and evaluation of various parameters related to the heart. For such evaluation, various cardiac structures can be contoured. For example, the left and right ventricles and left ventricular muscle structures can be delineated to calculate the volume of the cavities in diastole and systole (hence the ejection fraction) as well as the myocardial mass. These parameters can be useful for detecting and quantifying various pathologies. Manual tracing of these cardiac structures is time consuming due to the large amount of data involved, for example, due to the number of imaging slices acquired. For example, a cine MRI of a patient may have more than 20 phases and more than 10 short-axis slices per phase. It may take a trained professional approximately 30 minutes to manually trace the ventricles of a single subject and analyze the images.

[0003] An example clinical workflow uses semi-automatic techniques to segment the heart chambers in one phase of a cine cardiac MRI and propagate those contours to other phases. One major challenge is the lack of accurate automatic ventricular segmentation in the basal region of the heart. In some cases, experts may need to spend a significant amount of time manually correcting the segmentation output in the apical and basal slices of the cardiac MRI.

[0004] One of the main clinical parameters evaluated in cardiac MRI is the left ventricular ejection fraction. Ventricular segmentation can be used to determine the ejection fraction from the basal imaging slice. Automatic segmentation methods may struggle to distinguish between the right ventricle (RV), left ventricle (LV), right atrium (RA), left atrium (LA) and surrounding structures in the basal imaging slice. Manual editing of the automatic segmentation may be required to improve the accuracy of the segmentation, increasing the time spent by the segmentation expert before pathology can be assessed. Certain structures, such as those in the apical and basal image slices, may be more difficult to automatically segment than structures in the mid-ventricular image slices.

[0005] Techniques such as deep learning may provide accurate identification of strong boundaries, for example, when compared to model / atlas-based segmentation. However, segmentation in basal slices (and certain other anatomical features) may be difficult due to poor contrast (e.g., in terms of pixel intensity gradients) at the atrial / ventricular interface. Certain deep learning techniques may result in poor segmentation in apical / basal imaging slices and among various other anatomical features. Summary of the Invention [Problem to be solved by the invention]

[0006] A method, non-transitory machine-readable medium and apparatus for identifying anatomical features are provided.

[0007] Aspects or embodiments described herein may relate to improving the identification of specific anatomical features from imaging data. Aspects or embodiments described herein may avoid one or more problems associated with training machine learning models to identify specific anatomical features from imaging data. [Means for solving the problem]

[0008] In a first aspect, a method is described. The method is a computer-implemented method. The method includes receiving imaging data representing a volume of an anatomical structure of a subject. The received imaging data includes at least one unidentified anatomical feature of interest. The method further includes using a machine learning model configured to perform a segmentation technique for identifying at least one anatomical feature of interest in the received imaging data to identify the anatomical feature of interest in the received imaging data.

[0009] The machine learning model is trained using maps generated for each of a series of training data sets, each of which includes training imaging data representing a volume of an anatomical structure on which to train.

[0010] The map for each training data set is generated by a spatial function configured to specify a spatial distribution of at least one training region relative to at least one control location in the training data set associated with the map, the map being configured to penalize learning errors in the at least one training region, the at least one training region including at least one unidentified anatomical feature of interest in the training data set associated with the map. Several embodiments relating to the first and other aspects are described below.

[0011] In some embodiments, the machine learning model is configured to identify anatomical features of interest by: using a segmentation technique to determine where the identified anatomical features of interest are located in the imaging data; and generating an indicator indicating where the identified anatomical features of interest are located in the imaging data.

[0012] In some embodiments, at least one control location does not overlap with an anatomical feature of interest.

[0013] In some embodiments, using the machine learning model includes using an additional map as an input to the machine learning model to identify at least one anatomical feature of interest in the received imaging data, the additional map being generated from the received imaging data.

[0014] In some embodiments, the spatial overlap between adjacent training regions specified by the map generated for each training dataset defines at least one preferred training region in that training dataset for use by the machine learning model to prioritize penalization of training errors in the at least one preferred training region over non-overlapping training regions of the training dataset and / or penalization of training errors in other regions of the training dataset.

[0015] In some embodiments, the received imaging data corresponds to a base region of the subject's heart. The at least one anatomical feature of interest to be identified using the trained machine learning model may include at least one anatomical interface between adjacent cavities (lumens) of the subject's heart.

[0016] In some embodiments, the at least one control location is identified based on the results of an initial segmentation model used to identify the at least one control location.

[0017] In some embodiments, the at least one control location includes: a center of gravity of a chamber of the heart; and / or an end point and / or a junction of ventricular and / or atrial muscle structures that define at least one interface between the chambers of the heart.

[0018] In some embodiments, the spatial distribution of the at least one training region is defined by at least one parameter of a spatial function, which may be based on at least one dimension of at least one previously identified anatomical feature in the training data set.

[0019] In some embodiments, the spatial function includes a first Gaussian-based function centered on an origin defined by at least one control location in the training data set, and a spatial distribution of at least one training region defined by the first Gaussian-based function may be distal to the origin.

[0020] In some embodiments, the first Gaussian-based function includes an inverse Gaussian function.

[0021] In some embodiments, the volume includes at least a portion of a heart. The first Gaussian-based function may be centered on a center of gravity of at least one chamber of the heart.

[0022] In some embodiments, the spatial function includes a second Gaussian based function that specifies a spatial distribution indicative of at least one training region associated with the second Gaussian based function, the spatial distribution indicative of at least one training region associated with the second Gaussian based function overlapping adjacent control locations in the training data set.

[0023] In some embodiments, the volume includes at least a portion of a heart. The spatial distribution of the at least one training region defined by the second Gaussian-based function may include lines connecting adjacent endpoints and / or junctions of ventricular and / or atrial musculature that define at least one interface between chambers of the heart.

[0024] In some embodiments, a loss function used to penalize training errors is modified by the map, and may be based on a difference between a measurement value and a ground truth value for at least one pixel or voxel of the training imaging data.

[0025] In some embodiments, the method further comprises training a machine learning model. The method comprises receiving at least one of a series of training data sets and an indication of ground truth identifying an anatomical feature of interest in each training data set. The method further comprises determining at least one control position in the at least one training data set. The method further comprises generating a map for the at least one training data set by using a spatial function to generate a set of loss values. The set of loss values ​​is indicative of a spatial distribution of the at least one training region. The set of loss values ​​may be indicative of a loss function to apply to each pixel or voxel of the training data to penalize a learning error at each pixel or voxel. The method further comprises training a machine learning model using at least one of the series of training data sets and a corresponding map of the at least one training data set.

[0026] In a second aspect, a non-transitory machine-readable medium is described. The non-transitory machine-readable medium stores instructions executable by at least one processor. The instructions are configured to cause the at least one processor to receive imaging data representing a volume of an anatomical structure of a subject. The received imaging data includes at least one unidentified anatomical feature of interest. The instructions are further configured to cause the at least one processor to use a machine learning model configured to perform a segmentation technique for identifying the at least one anatomical feature of interest in the received imaging data to identify the anatomical feature of interest in the received imaging data. The machine learning model is as referred to in the first aspect and related embodiments.

[0027] In a third aspect, an apparatus is described. The apparatus includes at least one processor communicatively coupled to an interface. The interface is configured to receive imaging data representative of a volume of an anatomical structure of a subject. The received imaging data includes at least one unidentified anatomical feature of interest. The apparatus further includes a non-transitory machine-readable medium storing instructions readable and executable by the at least one processor. The instructions are configured to cause the at least one processor to use a machine learning model configured to perform a segmentation technique for identifying at least one anatomical feature of interest in the received imaging data to identify the anatomical feature of interest in the received imaging data. The machine learning model is as referenced in the first aspect and related embodiments.

[0028] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

[0029] Exemplary embodiments of the present invention will now be described, by way of example only, with reference to the following drawings, in which: [Brief description of the drawings]

[0030] [Figure 1] FIG. 1 illustrates a method for identifying anatomical features according to one embodiment. [Diagram 2] FIG. 2 is a schematic diagram of a system for identifying anatomical features according to one embodiment. [Diagram 3] FIG. 3 illustrates an image in which specific anatomical features have been identified according to one embodiment. [Figure 4] FIG. 4 is a flow chart illustrating a method for identifying anatomical features according to one embodiment. [Diagram 5] FIG. 5 illustrates an image in which specific control positions have been identified according to one embodiment. [Figure 6] FIG. 6 illustrates the spatial functions used in certain embodiments. [Figure 7] FIG. 7 illustrates other spatial functions that may be used in certain embodiments. [Figure 8] FIG. 8 illustrates a map generated based on an image in accordance with certain embodiments. [Figure 9] FIG. 9 illustrates a method for generating display data according to one embodiment. [Figure 10] FIG. 10 illustrates a method for training a machine learning model according to one embodiment. [Figure 11] FIG. 11 is a schematic diagram of a machine-readable medium for identifying anatomical features according to one embodiment. [Figure 12] FIG. 12 is a schematic diagram of an apparatus for identifying anatomical features according to one embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0031] A machine learning model (such as a deep neural network based on the “U-Net” architecture described in Ronneberger et al., “U-Net: Convolutional Networks for Biomedical Image Segmentation,” arXiv:1505.04597 (2015), the entire contents of which are incorporated herein by reference) can be trained to identify specific anatomical features through segmentation of imaging data acquired by radiological imaging devices such as MRI scanners, CT scanners, etc. In some exemplary training procedures, the machine learning model can attempt to reduce identification error by penalizing models that produce errors between the output from the model and “ground truth” data.

[0032] In these training procedures, each region of the imaging data used for training is uniformly penalized so that all regions of the imaging data can be treated equally during the training procedure. However, some regions in the imaging data may be difficult for machine learning models to accurately identify. Examples of difficult regions include interfaces between certain anatomical features because the intensity contrast of pixels / voxels spanning such interfaces is low / weak. While certain anatomical features / regions of the heart (such as the centroids of the chambers) can be accurately detected using certain machine learning based models (e.g., U-Net architecture) or certain "traditional" techniques (e.g., model / atlas based segmentation), certain interfaces (such as the atrium / ventricular interface) may be difficult to accurately detect with any of such techniques.

[0033] The framework described herein allows the machine learning model to better focus on learning to identify specific regions of the imaging data associated with poor / weak interfaces (i.e., where there may be poor / weak pixel / voxel intensity contrast across the interface). As described herein, the framework is described for cardiac MRI data. However, the framework can be generalized to other anatomical sites or structures such as the brain, liver, etc. Furthermore, the imaging data can be acquired using any suitable radiological imaging technique, such as CT scan, ultrasound-based imaging, etc.

[0034] 1 illustrates a method 100 for identifying anatomical features according to one embodiment. Method 100 is a computer-implemented method and may be implemented on a computing device as described below. For example, method 100 may be performed by a user computer (e.g., associated with use by a clinician such as a radiologist), a server, a cloud-based service, and / or any other computing device, whether at or off-site with a radiology imaging device for acquiring imaging data.

[0035] The method 100 includes receiving imaging data representative of a volume of an anatomical structure of a subject, at block 102. The received imaging data includes at least one unidentified "anatomical feature of interest."

[0036] A radiological imaging device can acquire at least one image (i.e., "imaging data") by imaging at least a portion of a subject's anatomy. A set of images (e.g., image slices) acquired by a radiological imaging device can be registered together and collectively represent a volume of the subject's anatomy (e.g., a portion of the body that includes at least one anatomical site of interest, such as the subject's heart). In some cases, a single set of images representing the volume may be acquired. In other cases, for example, when a portion of the body (e.g., the heart) moves, multiple images for each slice location may be acquired over a period of time.

[0037] There may be at least one unidentified anatomical feature of interest within the volume, such as an interface that appears as a "weak" difference in pixel / voxel intensity in that region. The unidentified anatomical feature of interest may either have been previously unidentifiable or may be a previously identified anatomical feature that cannot be verified to be a correctly identified anatomical feature of interest.

[0038] In the case of the heart, the unidentified anatomical feature of interest may be, for example, the atrium / ventricular interface as shown in the imaging data corresponding to the basal / apical regions. In some cases, at least one anatomical feature (other than the unidentified anatomical feature of interest) may already have been identified in the imaging data (e.g., based on an initial segmentation process using deep learning or conventional techniques). For example, in the case of the heart, the previously identified anatomical feature may be the centroid of the heart cavity (and the approximate shape and location of the feature), which may be easier for a particular segmentation method to identify than the atrium / ventricular interface or certain other interfaces.

[0039] At block 104, the method 100 includes identifying an anatomical feature of interest in the received imaging data using a machine learning model configured to perform a segmentation technique to identify at least one anatomical feature of interest in the received imaging data.

[0040] The segmentation technique, which may include a deep learning technique such as that based on the aforementioned "U-Net" architecture, may enable the machine learning model to identify at least one anatomical feature of interest within the received imaging data.

[0041] For example, the radiological imaging device may acquire imaging data of a subject, such as a patient. The computing device may receive the imaging data and then apply a segmentation technique according to block 104 to identify at least one anatomical feature of interest. Identifying the at least one anatomical feature of interest may include generating data representative of a segmentation of at least one region of the imaging data from other regions of the imaging data. The data thus generated may be used to visually represent the segmented portion (or portions) within the volume, e.g., to be displayed to indicate the segmented regions. A clinician or other operator may use the segmentation technique to visualize previously unidentified anatomical features of interest.

[0042] The machine learning model is trained using a map (e.g., a "penalty map" or "heat map") generated for each of a series of training data sets. Each training data set includes training imaging data representing a volume of an anatomical structure to be trained. The map is configured to cause the machine learning model to penalize learning errors in at least one training region that includes the at least one unidentified anatomical feature of interest in the training data set associated with the map. In other words, during the training procedure, the map can be used to penalize learning errors in regions of the training imaging data that include the at least one anatomical feature of interest. In some cases, the map can function as an "attention" mechanism to focus the model on at least one region of interest when training the model. Thus, in some cases, the map can be used to calculate a "loss function" for penalizing learning errors at at least one point (e.g., pixel or pixel region) in the training imaging data (as described below) and / or can function as an attention mechanism during training. Furthermore, if the map is used as an attention mechanism during training, such a map may be generated based on the received (input) imaging data and used as an additional input to the received (input) imaging data as an attention mechanism for focusing the model on at least one region of interest that contains at least one anatomical feature of interest when performing method 100.

[0043] In this manner, training may be based on a series of training data sets, where each training data set in the series includes imaging data corresponding to a part of a training subject's body (i.e., the training subject's body part may be referred to herein as a "training anatomical structure"). The part of the body includes the at least one anatomical feature of interest. For example, if the at least one anatomical feature of interest is in the heart, each training data set may include cardiac imaging data from a set of training subjects. At least one expert, such as a senior radiologist, may provide a "ground truth" that identifies the anatomical feature of interest in each training data set in the series.

[0044] By way of example, implementation of certain embodiments described herein is illustrated based on a set of data including 20 cardiac "training" data sets, where k-fold (e.g., 10-fold) cross-validation was used to test the accuracy of the trained model. The set of cardiac training data sets used included a selection of different cardiac image slices (i.e., from different patients at different cardiac phases). The model can be trained and tested on different numbers of the set of data sets. Furthermore, enhancement techniques such as adding Gaussian noise can be used to generate additional training data sets for the set. In this example involving 20 pediatric training data sets, a senior radiologist provided the "ground truth" locations of anatomical features of interest in each training data set. An exemplary process of training and testing a machine learning model based on 20 training data sets is described below. As described below, the training resulted in a model that provided accurate segmentation that may facilitate automated segmentation. Other training data sets can also be selected based on similar principles.

[0045] An exemplary training process that can be used is described in further detail below with reference to the process used in the present example (ie, training using a series of 20 pediatric "training" data sets).

[0046] The 20 pediatric cardiac "training" data sets used had a high degree of structural variability across the series. The four cardiac chambers and myocardium were manually segmented by a senior radiologist to represent the "ground truth" for each data set.

[0047] Difficult regions of interest considered by the model in the imaging data may be identified by an expert (ie, a senior radiologist) or may be automatically identified based on an error in at least one region of interest according to a particular baseline model.

[0048] Certain anatomical features such as centroids / major angles can aid in segmentation in difficult regions. In some cases, the magnitude of variation can be manually identified by an expert, allowing the location and structure (shape / size, etc.) of anatomical features to be explicitly learned by the model. However, as used in the training process of this example, feature structures can be estimated by training an initial model with a U-Net model (which may include a loss function) to obtain an initial segmentation. From that initial segmentation, potentially relevant feature structures can be automatically identified, again using image processing methods.

[0049] Based on the feature structures extracted by the initial segmentation, a "penalty" map was generated. The penalty map was then fed into the (2D sliced) U-Net model together with the input imaging data and ground truth expert segmentations for training. In this case, the penalty map was used in two ways: (1) to penalize training in difficult regions; and (2) to provide an "attention" mechanism that helps the model to focus on previously learned regions, thereby simplifying segmentation refinement after the initial segmentation.

[0050] The above example refers to the use of a set of 20 pediatric cardiac training data sets. The results of this experiment are described below. In another experiment, a set of 50 adult cardiac training data was used to train the machine learning model referenced in method 100. In the other experiment, the training method / attention mechanism described herein was validated by improving the segmentation of difficult feature structures of interest.

[0051] The generation of the map for the training data set is described with reference to method 100. The same principles of map generation can be applied when the map is generated based on received (input) imaging data.

[0052] The map for each training data set is generated by a spatial function configured to specify a spatial distribution of at least one training region for at least one control location in the training data set associated with the map, the map is configured to penalize learning errors in the at least one training region, and the at least one training region includes at least one unidentified anatomical feature of interest in the training data set associated with the map.

[0053] In certain embodiments, this penalization can be achieved by modifying a loss function implemented by the machine learning model during the training process. For example, the spatial function may generate a set of loss values. The set of loss values ​​may indicate a spatial distribution of at least one training region. The set of loss values ​​may indicate a loss function to apply to each pixel or voxel of the training dataset to penalize the learning error at each pixel or voxel. In other words, the machine learning model may use the loss function to determine where (i.e., at what pixels / voxels) in the training dataset to penalize losses (and how much to penalize the losses). The spatial distribution may indicate the associated pixels / voxels at which losses should be most / least penalized (where the loss value for each pixel / voxel indicates how much the losses should be penalized at each pixel / voxel). A set of loss values ​​corresponding to a spatial distribution may indicate that the loss is most penalized in regions having anatomical features of interest that are difficult to identify (i.e., at least one "training region"), and is less penalized in regions that are easier to identify. In this way, the machine learning model can focus attention (and efficiently allocate computational resources during the training process) on regions having anatomical features of interest that are difficult to identify.

[0054] In some cases, the at least one control location may correspond to at least one previously identified location in a training data set, for example, the at least one control location may correspond to a particular region of the training anatomical structure, such as a centroid of a heart chamber, that may in some cases be accurately identified from a segmentation technique based on a U-Net architecture or the like.

[0055] A map can be generated using a spatial function applied to at least one control location. Applying the spatial function to the at least one control location generates a spatial distribution defined by a map. The spatial distribution can define the location of at least one training region relative to the at least one control location. The spatial distribution can define a set of loss values ​​corresponding to how much loss should be penalized by each pixel / voxel represented by the training data set. By applying the spatial function / generating the spatial distribution, the resulting map can include at least one training region designated according to the spatial distribution associated with the spatial function. In other words, the map can indicate what regions of the training data set represent at least one "training region" derived from the spatial distribution. When using the map during training, the at least one training region can indicate what regions of the training imaging data should be focused on to learn (i.e., reduce error) regarding the at least one anatomical feature of interest. In some cases, the at least one training region can at least partially overlap with the at least one anatomical feature of interest. In this manner, determining the spatial distribution of at least one training region can provide an indication of where anatomical features of interest (e.g., difficult to identify) are found within the training imaging data, such that the machine learning model can be trained to reduce classification errors in such training regions.

[0056] Certain embodiments described herein may provide the ability to use a trained machine learning model to facilitate identification of a particular anatomical feature. For example, the at least one control position, in combination with application of a spatial function, may provide guidance to a segmentation model, such as one implemented by a U-Net architecture, as to where to focus its learning of the training imaging data to improve identification of at least one anatomical feature of interest.

[0057] The map resulting from the combination of at least one control position and application of the spatial function can define a "penalty" or "heat" map that indicates which regions of the training imaging data should be focused on learning in order to reduce identification errors. For example, in regions where there is the greatest "penalty" as indicated by the relative "heat" between different regions of the training imaging data, the machine learning model can attempt to learn by minimizing the error in such regions. In other words, the map can indicate to the model which regions in the training imaging data are useful for the machine learning model to better understand in order to identify at least one anatomical feature of interest.

[0058] For example, the map can facilitate penalizing automatic model learning based on distance from at least one control region, such that a maximum penalty is applied to at least one training region that includes the ventricular / atrial interface, the ventricular interface, and / or any other interface of interest.

[0059] Certain anatomical features that are difficult to identify accurately cannot be easily identified if each region of the training imaging data is treated equally, as in some techniques that may "blindly" learn, detect and / or segment based on spatial intensity patterns with "strong" (i.e., high pixel contrast) boundaries that are easier to detect. If each region of the training imaging data is treated equally, the resulting final classification error (which can be quantified by the "dice" score used to indicate the classification accuracy of the segmentation model) is an average over all regions. Thus, even if the overall "dice" score is high for the received imaging data, the accuracy of the segmentation in certain regions, such as in the atrium, may be low (i.e., the dice score may appear good overall, but this may hide classification errors in certain regions, such as the atrium region). However, the map generated can facilitate training improvements by increasing attention (e.g., by penalizing) given to certain regions of the training imaging data associated with anatomical features that are difficult to identify accurately.

[0060] The trained machine learning model was tested on the set of training datasets mentioned above (i.e., a set of 20 pediatric cardiac datasets), and the model provided segmentation with an accuracy of 0.931 "dice" score (compared to a baseline accuracy of 0.891) with a standard deviation of 0.018 (compared to a baseline standard deviation of 0.04). This improvement in accuracy demonstrated by the experimental data attests to the effectiveness of the approach described herein.

[0061] In some trained machine learning models for identifying anatomical features, such models may generate a probability map (indicating the probability that an anatomical feature is at a particular location in an image) as a mere output of the identification process (i.e., the map has no other purpose). However, the maps described herein in connection with various embodiments (which may or may not be the output of one / other trained machine learning model) may indicate where in the training data losses in a particular region of interest should be most penalized (i.e., for use during the training process itself). Such embodiments may use such a map generated for at least one control location identified in the training data. Such a control location may be easier to identify than the anatomical feature of interest itself (e.g., the centroid of a heart chamber may be relatively easier for one / other machine learning model to identify compared to the base region of the heart). The spatial function described herein can facilitate the machine learning model to identify where in the training data to focus its learning (e.g., the spatial function can define a spatial distribution (relative to a control location) of regions where losses are penalized more than other regions during the training process). As a result of using a map to penalize losses in certain regions (referred to herein as training regions), the machine learning model can be instructed to focus its learning on regions of the training data that are expected to contain anatomical features of interest (difficult to identify). By penalizing losses in difficult to identify regions, rather than penalizing losses in the same way for all regions in the training data, fewer computational resources can be used during the training process. In other words, using the map for the training process can improve the training of the model to identify difficult to identify features, while at the same time reducing or avoiding the need to allocate computational resources to overly penalizing losses in regions that are not that difficult to identify in the first place.

[0062] Accordingly, certain embodiments (e.g., method 100 and / or related embodiments) can facilitate improved accuracy for identifying at least one anatomical feature of interest within received imaging data. This improved accuracy can enable more accurate segmentation, and potentially automated segmentation with reduced or no manual input from an expert. This improved segmentation accuracy can be useful for certain views associated with anatomical features that are difficult to accurately identify, such as certain boundaries in the basal and / or apical regions of the heart.

[0063] The map may be computationally easy to generate and easily incorporated into a training procedure to facilitate automatic training / guidance of machine learning models. The map may provide an indication of which regions of the training imaging data to focus on, regardless of whether the training imaging data refers to one-dimensional, two-dimensional, three-dimensional, or four-dimensional imaging data. For example, the "map" may have any suitable dimension according to the format of the training imaging data, such that appropriate regions of the training imaging data are indicated as relevant for focused learning (e.g., by penalizing learning in such regions).

[0064] 2 illustrates an exemplary system 200 for implementing certain embodiments described herein (e.g., method 100 and certain other embodiments). System 200 includes a radiological imaging device 202, such as a magnetic resonance imaging (MRI) scanner, a computed tomography (CT) scanner, or the like. In use of system 200, a subject, such as a patient 204, and other medical equipment (not shown) are supported by a table 206 associated with radiological imaging device 202. Radiological imaging device 202 is communicatively coupled to a controller 208 (an example of what is referred to as a "computing device" in certain embodiments) for transmitting / receiving data to / from radiological imaging device 202, such as control data for controlling / monitoring the operation of radiological imaging device 202 and / or imaging data acquired by radiological imaging device 202. Controller 208 is communicatively coupled to a user interface, such as a display 210, for displaying imaging data and / or other information related to use of system 200.

[0065] In some cases, the controller 208 may be communicatively coupled to an optional service provider 212 (e.g., a manufacturer of the radiation imaging device 202 or other entity that may control / monitor / perform data processing in connection with the radiation imaging device 202 and / or the controller 208) as shown in FIG. 2. The service provider 212 may be a server or cloud-based service to which the controller 208 may be connected to exchange various data. In some cases, the service provider 212 may provide data processing services (e.g., to at least partially implement certain embodiments described herein). In some cases, the controller 208 may perform data processing (e.g., to at least partially implement certain embodiments described herein). In some cases, the controller 208 and the service provider 212 may exchange data and / or jointly perform / facilitate data processing to implement certain embodiments described herein. In some cases, the controller 208 may receive updates (e.g., software updates, etc.) from the service provider 212. Such updates may include information regarding new and / or updated models (e.g., a trained machine learning model and / or parameters, such as neural network weights, for the controller 208 to implement such trained machine learning model). Thus, in one scenario, the service provider 212 may train a machine learning model (e.g., according to certain embodiments described herein) and then transmit the trained machine learning model (or information, such as parameters for enabling such machine learning) to the controller 208 (to enable the controller 208 to implement the trained machine learning model according to certain embodiments described herein). In other scenarios, the controller 208 may be pre-loaded with the trained machine learning model.In other scenarios, the controller 208 may not implement the machine learning model, but instead the service provider 212 may implement a trained machine learning model (e.g., based on imaging data sent by the controller 208 to the service provider 212). In this manner, the controller 208 and / or the service provider 212 may implement a "computing device" as referenced in various embodiments described herein (e.g., method 100 and other embodiments described herein).

[0066] Controller 208 and service provider 212 (if present) may each include processing circuitry (such as at least one processor, not shown) configured to perform data processing to implement certain embodiments described herein. Controller 208 and / or service provider 212 include or have access to memory (e.g., a non-transitory machine-readable medium) that stores instructions that, when executed by the processing circuitry, cause the processing circuitry to implement certain embodiments described herein.

[0067] In some cases, the controller 208 can be implemented by a user computer. In some cases, the controller 208 and / or the service provider 212 can be implemented by a server or a cloud-based computing service. In some cases, memory (such as the non-transitory machine-readable media described above, and / or other memory, such as other non-transitory or transitory machine-readable media) can store information related to the machine learning model (e.g., the machine learning model itself and / or parameters associated with the model) and / or other data, such as imaging data associated with the radiology imaging device 202.

[0068] FIG. 3 illustrates imaging data including images including a base region of the heart (a) before and (b) after identification of anatomical features according to certain embodiments described herein. Image (a) may have been acquired by radiology imaging device 202 of FIG. 2. In this case, controller 208 and / or service provider 212 may receive the imaging data and automatically identify at least one anatomical feature of interest according to certain embodiments described herein (e.g., method 100). In this case, image (b) illustrates this identification by showing a segmentation (of white lines) between the cavities of the heart. As shown in FIG. 3, the segmentation achieved according to embodiments described herein overcomes the problem of low pixel contrast across certain boundaries (e.g., the atrial / ventricular boundary) as best illustrated by image (a).

[0069] 4 is a flowchart 400 outlining system operations (e.g., to be performed by at least a portion of system 200, such as controller 208 and / or service provider 212) for implementing certain embodiments described herein (e.g., method 100 and / or other embodiments). Various blocks are illustrated in FIG 4, and certain blocks may be omitted according to certain embodiments.

[0070] In some cases, such model training may be performed prior to implementing certain embodiments described herein (e.g., the "received imaging data" may not contribute to training). In some cases, some model training may be performed as part of implementing certain embodiments described herein (e.g., as seen in FIG. 4, the "received imaging data" may be used for further model training). In some cases, the level of expert input may vary. For example, manual correction may or may not be required.

[0071] In some cases, it may be necessary to generate a new penalty map (e.g., an "additional map" as described herein) for each group of "received imaging data" to provide guidance (i.e., as an attention mechanism) used by the machine learning model to identify at least one anatomical feature of interest from the received imaging data. In some cases, a penalty map may not need to be generated for each received imaging data (e.g., when improvements in the training of the machine learning model allow the machine learning model to perform discrimination independently based only on the received imaging data). In other cases, a new penalty map does not need to be generated for the model at deployment time, for example, when model training is performed using a penalty map only in the loss function. Thus, if a penalty map is used as an attention mechanism during training, a new penalty map may be used when the model is deployed to segment received imaging data.

[0072] In block 402 of flowchart 400, imaging data, such as at least one image slice, is received (eg, similar to block 102 of method 100).

[0073] At block 404 of flowchart 400, at least one control position (as referenced in method 100) is detected within the received imaging data.

[0074] In block 406 of flowchart 400, a map is generated (as referenced in method 100).

[0075] The flowchart 400 proceeds to at least one of blocks 408, 410, and / or 412, depending on the configuration.

[0076] At block 408, a penalty dependent loss is calculated based on the map to determine how much to penalize errors in different regions of the imaging data.

[0077] In block 410, the machine learning model is trained based on the map generated in block 406 or based on the penalty-dependent loss calculated in block 408.

[0078] In block 412, a machine learning model (either trained in block 410 or previously trained) is run based on the received imaging data (in block 402) and the corresponding map (in block 406) to identify at least one anatomical feature of interest (e.g., similar to block 104 of method 100).

[0079] In block 414, the results of running the model in block 412 may be corrected (e.g., manually by a user, such as an expert). This correction may be fed back to the control position detection in block 404.

[0080] The following description will refer to at least some possible configurations of the blocks of flowchart 400 in the context of identifying a boundary surface within received imaging data representing a base region of the heart. Reference will also be made to features of the previous figure.

[0081] Block 404 of flow chart 400 refers to control position detection. Based on input Cardiac MRI (CMRI) images (described below), a set of control positions is defined. In this configuration, the set of control positions includes the centers of gravity of the LV, RV, LA, and RA, as well as two end points (or "junctions") that represent the origin of the ventricular musculature.

[0082] The centers of gravity of the LV, RV, LA, and RA can be obtained as follows:

[0083] An initial segmentation model (e.g., any segmentation method such as one based on the U-Net architecture) is trained to obtain an initial segmentation of the LV, RV, LA, and RA.

[0084] From the 4-chamber segmentation, the corresponding centroids of the LV, RV, LA and RA are obtained (i.e., center_LV, center_RV, center_LA, center_RA, respectively). The hypothesis is that the centroids of the initial segmentation are relatively accurate even if the overall Dice score is low.

[0085] Further, the size of each structure (e.g., at least one internal dimension of LV, RV, LA and RA in each image slice) is also estimated by identifying at least one direction and magnitude of change corresponding to the size of the structures (and their magnitude of change), such as their lengths in the first two principal directions, from the initial model using a technique such as principal component analysis (PCA) on the segmented regions of LV, RV, LA and RA to obtain the length of each cavity, i.e., length_LV, length_RV, length_LA, length_RA.

[0086] The endpoints of the LV myocardial structure (referred to herein as "EP1" and "EP2") can be detected by obtaining an initial myocardial segmentation (e.g., using "traditional" non-machine learning based image processing and / or machine learning methods), and then EP1 and EP2 are identified by finding the origin of the ventricular myocardial structure (e.g., in the long axis 4 chamber view) so that the myocardium can be segmented. The points detected on either side of the LV where the ventricular myocardium is thinnest represent the endpoints of each slice.

[0087] 5 shows points "EP1" and "EP2" in an example image 500 illustrating the initial segmentation and myocardial segmentation described above. A dotted line 502 is shown connecting EP1 and EP2, representing their approximate location in the base of the heart region where segmentation has not yet been determined.

[0088] A map is then generated based on the six control positions, as described below.

[0089] A map of the same size as the input image is initialized with zeros.

[0090] An inverse Gaussian function (which is an example of a “spatial function” referred to in method 100) is applied to each of the centroids (i.e., center_LV, center_RV, center_LA, center_RA), where each inverse Gaussian function has at least one standard deviation proportional to the corresponding length as described above.

[0091] FIG. 6 shows an example of an inverse Gaussian function 600 having a spatial distribution that produces a minimum value ("0") at the origin 602 (i.e., to be centered at each centroid location) and a maximum value ("1") at a specified distance from the origin (the specified distance may depend on the radial direction from the origin). Keys are provided to indicate the minimum and maximum values ​​in the map, as well as values ​​therebetween. The parameters used to generate the inverse Gaussian function 600 depend on at least one dimension of the cavity to which the inverse Gaussian function 600 is applied. For example, as shown by FIG. 6, the shape of the inverse Gaussian function 600 has major and minor dimensions 604 and 606, respectively, for a cavity having an approximately elliptical shape (i.e., the first dimension 604 is the maximum distance between opposite sides of the elliptical shape, and the second dimension 606 is the minimum distance between opposite sides of the elliptical shape). In this manner, the parameters for generating a centered inverse Gaussian in each cavity can be appropriately determined from at least one dimension 604, 606 of that cavity (as could be determined from the initial segmentation or by using other techniques).

[0092] A Gaussian "beam" function (which is another example of a "spatial function" referred to in method 100) is applied to superimpose a line 502 between points EP1 and EP2. The parameters used to generate the Gaussian beam function may depend on predefined values ​​based on the distance between EP1 and EP2 and the expected thickness of the anatomical feature of interest.

[0093] Figure 7 shows an example of a Gaussian beam function 700. The Gaussian beam function 700 shown in Figure 7 is overlapped with points EP1 and EP2. A first parameter defining a length 702 of the Gaussian beam function 700 may be selected based on the distance between EP1 and EP2 (i.e., the length 702 is selected to be longer than the distance to ensure the overlap). A second parameter defining a width 704 of the Gaussian beam function 700 (e.g., standard deviation) may be selected based on the expected thickness of the anatomical feature of interest.

[0094] A combination of the inverse Gaussian function referenced in FIG. 6 and the Gaussian beam function referenced in FIG. 7 can be used to generate a map that defines the penalty at pixel locations within the imaging data.

[0095] 8 shows (a) an image slice of a portion of the heart, and (b) a "heat" map generated based on image (a), where heat map (b) can be overlaid with image slice (a) to identify a region (or regions) of interest. As shown by heat map (b), there are four inverse Gaussian functions (with origin at the centroid), each of which at least partially overlaps (in the "brightest" regions of the heat map) with at least one other inverse Gaussian function and with a Gaussian beam function.

[0096] The maxima in the heat map correspond to the areas where the Gaussian functions associated with the various control locations overlap or intersect. In this case, the maxima correspond to the overlap of the Gaussian functions defined for the control locations, e.g., the area around the upper border of the septum and the basal slice, followed by the interface between the LV / LA and RV / RA, then other interfaces or points of each structure, and finally, the minima in the map correspond to the location of the centroid of the LV / RV / LA / RA and outside these structures. A caption is provided to indicate the maximum ("1") and minimum ("0") values ​​of the heat map. A maximum of "1" means the overlap of at least two different Gaussian functions, but the selection of the minimum and maximum values ​​is arbitrary depending on the design of the map and the training system.

[0097] The generated map can provide guidance to the model to identify the most difficult regions, allowing the model to automatically learn feature structures to optimize segmentation in these regions based on location and weighting.

[0098] The above-mentioned maps are generated based on certain spatial functions that have a spatial distribution (which may or may not overlap) that corresponds to the regions in which the machine learning model is going to penalize errors. Other spatial functions can be used if appropriate. Furthermore, while an inverse Gaussian function has been described, it is also possible to use a positive Gaussian function with origin at the centroid to point to the regions in which the machine learning model should not concentrate. In other words, the maps can be generated (and optionally transformed) in such a way as to force the machine learning model to focus its learning on certain regions of interest (including at least one anatomical feature of interest), whether this is performed by penalizing or in some other way that forces the machine learning model to prioritize learning on these regions of interest. Thus, it is not important whether the values ​​of the maps are positive or negative, have a large or small scaling range, etc., as long as they are appropriately transformed for use during training.

[0099] Block 410 of flowchart 400 refers to training the model (eg, using the "penalty" map described above).

[0100] During training of the model, a penalty map can be used to penalize the model to encourage it to produce accurate segmentations in difficult regions. In one configuration, this penalization can be performed by adding an extra penalty term to any suitable loss function used to train the segmentation model. By way of example, the loss function can be:

number

[0101] where Loss(i) is the loss for a given voxel i and I± are weighting coefficients associated with the penalty map, which can be determined experimentally. PenaltyMap(i) is the associated penalty at voxel i. This can be applied to a variety of loss functions. Below we show an example application to the (weighted / modified) Dice loss, which is widely used for segmentation problems. Thus, the modified Dice loss is:

number

[0102] where y j,l (y_j, l) and y' j,l (y'_j,l) are the ground truth and predicted labels of the jth pixel in the lth class, respectively, and w l(w_l) is the weight associated with class l. Note that the Dice value (equal to 1 means high segmentation accuracy, and equal to 0 means low segmentation accuracy) refers to the second term in the brackets. The structure of the Dice loss formula means that the Dice loss value will be a large number if the Dice score is low (i.e., poor segmentation accuracy at a particular pixel location, such as at a weak interface), and a small number if the Dice score is high (i.e., high segmentation accuracy at a particular pixel location, such as at a heart chamber centroid). Thus, the Dice loss for the jth pixel is used to indicate to the model whether to focus learning on a particular pixel or group of pixels (e.g., because the error at such pixel / group of pixels is large).

[0103] Therefore, if the segmentation accuracy is low in a region with high penalty, the model will suffer high loss and the model will automatically learn to minimize the segmentation error in this region, thus resulting in higher segmentation accuracy in a particular region, such as at least one region containing at least one anatomical feature of interest.

[0104] Block 412 of flowchart 400 refers to running the trained model. Once training is complete, the required control positions may, in some cases, be automatically generated and provided to the user for acceptance / modification. Once the user accepts the identified regions, a penalty map is generated and provided as input along with the imaging data to guide the segmentation.

[0105] Several embodiments related to the above are described below.

[0106] In some embodiments, the machine learning model is configured to identify anatomical features of interest by: determining where within the imaging data the identified anatomical features of interest are located using a segmentation technique; and generating an indicator to indicate where within the imaging data the identified anatomical features of interest are located. In some cases, the indicator may be configured to provide a visual indication (e.g., via a marker such as an arrow or text) of the identified anatomical features of interest on a representation (e.g., on a user display) of the imaging data.

[0107] In some embodiments, at least one control location does not overlap with an anatomical feature of interest. For example, the at least control location may be a location / anatomical feature that is easy for some / other machine learning model to identify. This control location may then be used to determine where to focus learning in the training dataset (e.g., via a spatial distribution defined for the at least one control location). In some cases, at least one of the control locations may overlap with an anatomical feature of interest (although at least one other of these control locations does not overlap with the anatomical feature of interest).

[0108] In some embodiments, using the machine learning model includes using an additional map as an input to the machine learning model to identify at least one anatomical feature of interest in the received imaging data. The additional map can be generated from the received imaging data. In other words, the additional map can be an additional input with the received imaging data. That is, the machine learning model actually processes the imaging data in addition to the additional map to extract features (i.e., perform identification of anatomical features of interest). It should be noted that the additional map generated from the received imaging data is separate from the map generated from each training data set. Thus, the use of the additional map is in addition to the same map being used to penalize the loss during the training process (e.g., by changing the loss function).

[0109] The additional map can be generated in a similar / same manner as the map used to penalize the loss during the training process. For example, the machine learning model (or other machine learning model) can determine at least one control location in the received imaging data (e.g., by identifying a relatively easy to identify location such as the cavity centroid). The machine learning model can then generate an additional map for the received imaging data by using a spatial function to generate a spatial distribution indicating where the machine learning model should focus its attention to identify anatomical features of interest (i.e., in a similar manner as the spatial function was used to generate the set of loss values). In this manner, the map is generated based on the training data set and used for the training process. Additional maps are generated based on the received imaging data and used for the identification process. These and additional maps can be represented as heat maps, attention maps, and / or penalty maps. In some cases, such as when the machine learning model uses a (training) map to help train the machine learning model, additional maps may be required as input to the identification process.

[0110] In some embodiments, the spatial overlap between adjacent training regions specified by the map generated for each training dataset defines at least one preferred training region in that training dataset that is used by the machine learning model to prioritize penalization of training errors in the at least one preferred training region over penalization of training errors in non-overlapping training regions of that training dataset and / or other regions of that training dataset.

[0111] As shown by FIG. 8b), the “brightest” regions of the map correspond to overlaps of different spatial functions (i.e., overlaps between spatial distributions defined by an inverse Gaussian function and a Gaussian beam function) that define multiple training regions (with different priorities) of the imaging data. Any spatial region that corresponds to a “spatial overlap” can define a highest priority region of the imaging data to be trained (e.g., at least one “high” priority “priority training region”), since the overlapping region may have the highest brightness. Spatial regions that correspond to non-overlapping spatial distributions of the spatial functions (i.e., the “non-overlapping training regions” discussed above) may define the next highest priority of the imaging data to be trained. Other regions of the map (i.e., the “other regions” discussed above) (i.e., the “darkest” regions) may define the lowest priority of the imaging data to be trained. In this manner, the lowest priority regions may indicate regions that are relatively easy for a segmentation model to identify in the imaging data, while the highest priority regions may refer to regions that are difficult for a segmentation model to identify in the imaging data without using the map described herein.

[0112] In some embodiments, the received imaging data corresponds to a base region of the subject's heart, in which case the at least one anatomical feature of interest to be identified using the trained machine learning model may include at least one anatomical interface between adjacent chambers of the subject's heart.

[0113] In some embodiments, the at least one control location is identified based on the results of an initial segmentation model used to identify the at least one control location.

[0114] In some embodiments, the at least one control location includes a center of gravity of a chamber of the heart. In some embodiments, the at least one control location includes an edge and / or junction of ventricular and / or atrial muscular structures that define at least one interface between the chambers of the heart. In some embodiments, the control location may include a combination of at least one center of gravity of a chamber of the heart and at least one edge and / or junction of ventricular and / or atrial muscular structures.

[0115] In some embodiments, the spatial distribution of the at least one training region is defined by at least one parameter of a spatial function. The at least one parameter may be based on at least one dimension of at least one previously identified anatomical feature in the training data set. As mentioned in connection with Figures 6 and 7, the dimension of the at least one anatomical feature (e.g., at least one dimension of a heart chamber) may define the spatial extent / distribution of the spatial function. Thus, the at least one parameter of the spatial function may be selected based on at least one dimension (measured or expected), such as the length and / or width, of the heart chamber in the image slice.

[0116] In some embodiments, the spatial function comprises a first Gaussian-based function centered on an origin defined by at least one control position in the training data set. The spatial distribution of at least one training region defined by the first Gaussian-based function may be distal to the origin. In other words, there may be a certain distance (which may vary according to the radial direction from the origin) between the spatial distribution of at least one training region defined by the first Gaussian-based function, so that the machine learning model does not focus on the origin, but focuses on the region defined by the spatial distribution.

[0117] In some embodiments, the first Gaussian-based function includes an inverse Gaussian function.

[0118] In some embodiments, the volume includes at least a portion of a heart, in which case the first Gaussian-based function may be centered on the centroid of at least one chamber of the heart.

[0119] In some embodiments, the spatial function includes a second Gaussian base function that specifies a spatial distribution, the spatial distribution being indicative of at least one training region associated with the second Gaussian base function. The spatial distribution indicative of at least one training region associated with the second Gaussian base function may overlap adjacent control positions in the training data set. This is in contrast to the first Gaussian base function, whose spatial distribution defining the region of interest does not overlap with the control positions (i.e., the origin).

[0120] In some embodiments, the volume includes at least a portion of a heart, where the spatial distribution of the at least one training region defined by the second Gaussian-based function may include lines connecting adjacent endpoints and / or junctions of ventricular and / or atrial musculature that define at least one interface between respective chambers of the heart.

[0121] In some embodiments, a loss function used to penalize training errors is modified by the map. In some embodiments, the loss function is based on the difference between a measurement value for at least one pixel or voxel of the training imaging data and a ground truth value. For example, the Dice loss described above is one example of a "loss function" that can be used to facilitate training of a machine learning model.

[0122] In some embodiments, the machine learning model may be initially trained based on a "training data set." In some embodiments, the machine learning model may be expanded based on this initial training. In some embodiments, the machine learning model may be further trained (e.g., to improve the machine learning model) using the "received imaging data" referenced in method 100. Thus, in some embodiments, the "training imaging data" includes the "received imaging data." In other embodiments, the "received imaging data" referenced in method 100 may not be used for further training.

[0123] 9 illustrates a method 900 for generating representation data according to one embodiment. Method 900 may be implemented on a computing device, such as those described above, similar to method 100, for example. In some embodiments, method 900 may be implemented independently of method 100. In this embodiment, method 900 is implemented in conjunction with method 100 (e.g., method 900 may be implemented after identification has been performed according to method 100).

[0124] Block 902 of method 900 includes generating display data for displaying a segmentation of the identified anatomical feature of interest relative to at least one other anatomical structure within the received imaging data on a user interface (e.g., display 210).

[0125] 10 illustrates a method 1000 for training a machine learning model according to one embodiment. Method 1000 may be implemented on a computing device, such as those described above, similar to method 100, for example. In some embodiments, method 1000 may be implemented independently of method 100. In this embodiment, method 1000 is implemented in conjunction with method 100 (e.g., method 1000 may be implemented to facilitate training of the machine learning model referenced in FIG. 1).

[0126] The method 1000 includes training a machine learning model as follows.

[0127] Method 1000 includes receiving at least one series of training data sets (e.g., each data set including at least one image slice of received imaging data, etc., and / or a previously acquired training data set) and ground truth indications identifying anatomical features of interest within each training data set, at block 1002. In some cases, block 1002 may be the same as or similar to block 402 of flowchart 400. The at least one image slice may be a two-dimensional (2D) image that can be used to construct a three-dimensional (3D) image by registering a series of 2D images together.

[0128] Method 1000 includes determining at least one control position in the at least one training data set, at block 1004. In some cases, block 1004 may be the same as or similar to block 404 of flowchart 400.

[0129] Method 1000 includes, at block 1006, generating a map for at least one training data set by generating a set of loss values ​​using (at least one) spatial function. The set of loss values ​​may indicate a spatial distribution of at least one training region. The set of loss values ​​may indicate a loss function to apply to each pixel or voxel of the training data set to penalize a learning error at each pixel or voxel. In some cases, block 1006 may be the same or similar to block 406 of flowchart 400.

[0130] Method 1000 includes training a machine learning model using at least one training data set of the series and a corresponding map of the at least one training data set at block 1008. In some cases, block 1008 may be the same as or similar to block 410 of flowchart 400. As previously discussed, the map may be used to generate a loss function and / or as an attention mechanism.

[0131] 11 illustrates a non-transitory machine-readable medium 1100 for identifying anatomical features according to one embodiment. The non-transitory machine-readable medium 1100 has instructions 1102 that, when executed on at least one processor 1104, cause the at least one processor 1104 to perform certain methods described herein (e.g., methods 100, 900, 1000 and / or any related embodiments). In this embodiment, the instructions 1102 are configured to perform method 100. The non-transitory machine-readable medium 1100 may be located within the controller 208 and / or the service provider 212 of FIG. 2. Additionally, the at least one processor 1104 may also be located within the controller 208 and / or the service provider 212 of FIG. 2. In this manner, the system 200 may be used to execute the instructions 1102, including the instructions described below.

[0132] The instructions 1102 include instructions 1106 configured to cause at least one processor 1104 to receive imaging data representative of a volume of an anatomical structure of a subject. The received imaging data includes at least one unidentified anatomical feature of interest.

[0133] The instructions 1102 further include instructions 1108 configured to cause the at least one processor 1104 to perform a segmentation technique to identify anatomical features of interest within the received imaging data and to use a machine learning model configured to identify anatomical features of interest within the imaging data.

[0134] The machine learning model is trained using maps generated for each of a series of training data sets, each of which includes training imaging data representing a volume of an anatomical structure on which to train.

[0135] The map for each training data set is generated by a spatial function configured to specify a spatial distribution of at least one training region relative to at least one control location in the training data set associated with the map, the map being configured to penalize learning errors in the at least one training region, the at least one training region including at least one unidentified anatomical feature of interest in the training data set associated with the map.

[0136] In some embodiments, the instructions 1102 include other instructions for performing any of the other methods described herein.

[0137] FIG. 12 illustrates an apparatus 1200 for identifying anatomical features according to an embodiment. The apparatus 1200 includes at least one processor 1202 (e.g., implemented by a computing device such as the controller 208 and / or the service provider 212 illustrated in FIG. 2). The at least one processor 1202 is communicatively coupled to an interface 1204 for communicating data (e.g., with the radiographic device 202 and / or any other entity with which the at least one processor 1202 may exchange data when in use). In this embodiment, the interface 1204 is configured to receive imaging data representative of a volume of an anatomical structure of a subject. The received imaging data includes at least one unidentified anatomical feature of interest. The interface 1204 may be part of the controller 208 and / or the service provider 212 referenced in FIG. 2.

[0138] The apparatus 1200 further includes a non-transitory machine-readable medium 1206 having stored thereon instructions 1208 readable and executable by the at least one processor 1202 to perform methods corresponding to particular methods described herein (e.g., methods 100, 900, 1000 and / or any other methods described herein).

[0139] The instructions 1208 are configured to cause the at least one processor 1202 to perform a segmentation technique to identify anatomical features of interest within the received imaging data and to use a machine learning model configured to identify anatomical features of interest within the received imaging data.

[0140] The machine learning model is trained using maps generated for each of a series of training data sets, each of which includes training imaging data representing a volume of an anatomical structure on which to train.

[0141] The map for each training data set is generated by a spatial function configured to specify a spatial distribution of at least one training region relative to at least one control location in the training data set associated with the map, the map being configured to penalize learning errors in the at least one training region, the at least one training region including at least one unidentified anatomical feature of interest in the training data set associated with the map.

[0142] In some embodiments, instructions 1208 may include other instructions for performing any of the other methods described herein.

[0143] This disclosure includes subject matter defined in the following numbered paragraphs: 1. A computer-implemented method comprising: - receiving imaging data representative of a volume of an anatomical structure of a subject, the received imaging data including at least one unidentified anatomical feature of interest; and - identifying an anatomical feature of interest within the received imaging data using a machine learning model configured to perform a segmentation technique to identify at least one anatomical feature of interest; having - the machine learning model is trained using a map generated for each of a series of training data sets, each training data set including training imaging data representing a volume of an anatomical structure to train on, the map being configured to cause the machine learning model to penalize learning errors in at least one training region that includes at least one unidentified anatomical feature of interest in the training data set associated with the map; - a map for each training data set is generated by applying a spatial function to at least one control location in the training data set associated with the map; and the spatial function is configured to specify a spatial distribution of at least one training region relative to at least one control location in a training data set associated with said map. 2. The method of paragraph 1, wherein the spatial overlap between adjacent training regions specified by the map generated for each training dataset defines at least one preferred training region in the training dataset for use by the machine learning model to prioritize penalization of learning errors in the at least one preferred training region over non-overlapping training regions of the training dataset and / or penalization of learning errors in other regions of the training dataset. 3. The method of paragraph 1. or paragraph 2., wherein the received imaging data corresponds to a base region of the subject's heart, and the at least one anatomical feature of interest to be identified using the trained machine learning model includes at least one anatomical interface between adjacent cavities of the subject's heart. 4. The method of any one of paragraphs 1 to 3, wherein the at least one control location is identified based on a result of an initial segmentation model used to identify the at least one control location. 5. The method of paragraph 4., wherein the at least one control location includes: a center of gravity of a chamber of the heart; and / or an end point and / or a junction of ventricular and / or atrial muscular structures that define at least one interface between the chambers of the heart. 6. The method of any one of paragraphs 1 to 5, wherein the spatial distribution of the at least one training region is defined by at least one parameter of a spatial function, the at least one parameter being based on at least one dimension of at least one previously identified anatomical feature in the training data set. 7. The method of any one of paragraphs 1 to 6, wherein the spatial function includes a first Gaussian-based function centered on an origin defined by at least one control location in the training data set, and the spatial distribution of at least one training region defined by the first Gaussian-based function is distal to the origin. 8. The method of paragraph 7., wherein the first Gaussian-based function includes an inverse Gaussian function. 9. The method of paragraphs 7. or 8., wherein the volume includes at least a portion of the heart and the first Gaussian-based function is centered on a centroid of at least one chamber of the heart. 10. The method of any one of paragraphs 1 to 9, wherein the spatial function includes a second Gaussian-based function that spans between adjacent control locations in the training data set, and a spatial distribution of at least one training region defined by the second Gaussian-based function includes the adjacent control locations. 11. The method of paragraph 10, wherein the volume includes at least a portion of the heart, and the spatial distribution of the at least one training region defined by the second Gaussian-based function includes lines connecting adjacent endpoints and / or junctions of ventricular and / or atrial musculature that define at least one interface between each chamber of the heart. 12. The method of any one of paragraphs 1 to 11, wherein a loss function used to penalize training errors is modified by the map, and the loss function is based on a difference between a measurement value and a ground truth value for at least one pixel or voxel of the training imaging data. 13. The method of any one of paragraphs 1 to 12, comprising: - receiving at least one of a set of training data sets and an indication of a ground truth that identifies an anatomical feature of interest within each training data set; - determining at least one control position in at least one training data set; - generating a map for at least one training data set by applying a spatial function to at least one control location in the training data set, where the map is indicative of a loss function to apply to each pixel or voxel of the training data set; and - training the machine learning model using at least one training dataset of the set of training datasets and a corresponding map for the at least one training dataset; The method further comprises a step of training the subject by performing the training. 14. A non-transitory machine-readable medium storing instructions executable by at least one processor, the instructions causing the at least one processor to: - receiving imaging data representative of a volume of an anatomical structure of a subject, the received imaging data including at least one unidentified anatomical feature of interest; and - identifying an anatomical feature of interest within the received imaging data using a machine learning model configured to perform a segmentation technique to identify at least one anatomical feature of interest; It is configured as follows: - the machine learning model is trained using a map generated for each of a series of training data sets, each training data set including training imaging data representing a volume of an anatomical structure to train on, the map being configured to cause the machine learning model to penalize learning errors in at least one training region including at least one unidentified anatomical feature of interest in the training data set associated with the map; - a map for each training data set is generated by applying a spatial function to at least one control location in the training data set associated with the map; and the spatial function is configured to specify a spatial distribution of at least one training region relative to at least one control location in a training data set associated with said map. 15. An apparatus comprising: - at least one processor communicatively coupled to the interface, the interface configured to receive imaging data representative of a volume of an anatomical structure of the subject, the received imaging data including at least one unidentified anatomical feature of interest; and - a non-transitory machine-readable medium storing instructions readable and executable by at least one processor, the instructions configured to cause the at least one processor to identify an anatomical feature of interest in the received imaging data using a machine learning model configured to perform a segmentation technique to identify at least one anatomical feature of interest; having - the machine learning model is trained using a map generated for each of a series of training data sets, each training data set including training imaging data representing a volume of an anatomical structure to train on, the map being configured to cause the machine learning model to penalize learning errors in at least one training region including at least one unidentified anatomical feature of interest in the training data set associated with the map; - a map for each training data set is generated by applying a spatial function to at least one control location in the training data set associated with the map; and the spatial function is configured to specify a spatial distribution of at least one training region relative to at least one control location in a training data set associated with said map.

[0144] While the invention has been illustrated and described in detail in the drawings and foregoing description, such illustrations and description are to be considered illustrative or exemplary and not restrictive, i.e., the invention is not limited to the disclosed embodiments.

[0145] One or more features described in one embodiment may be combined with, or substituted for, features described in other embodiments.

[0146] Embodiments of the present disclosure may be provided as a method, a system, or a combination of machine-readable instructions and processing circuitry. Such machine-readable instructions may be included in a non-transitory machine (e.g., computer) readable storage medium (including, but not limited to, disk storage, CD-ROM, optical storage, flash storage, etc.) having computer-readable program code.

[0147] The present disclosure is described with reference to flowcharts and block diagrams of methods, apparatuses and systems according to embodiments of the present disclosure. Although the flowcharts described above show a specific order of execution, the order of execution may differ from that shown. Blocks described in relation to one flowchart may be combined with those of other flowcharts. It should be understood that each block in the flowcharts and / or block diagrams, and combinations of blocks in the flowcharts and / or block diagrams, may be implemented by machine-readable instructions.

[0148] The machine-readable instructions can be executed, for example, by a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to realize the functions described in the description and drawings. In particular, a processor or processing circuit, or modules thereof, can execute the machine-readable instructions. Thus, the functional modules of the apparatus and other devices described herein can be embodied by a processor that executes machine-readable instructions stored in a memory, or a processor that operates according to instructions embedded in a logic circuit. The term "processor" should be interpreted broadly to include a CPU, a processing unit, an ASIC, a logic unit, or a programmable gate array, etc. The methods and functional modules can all be executed by a single processor, or can be divided among several processors.

[0149] Such machine-readable instructions may be stored in a computer-readable storage device that causes a computer or other programmable data processing apparatus to operate in a particular mode.

[0150] Such machine-readable instructions may be loaded into a computer or other programmable data processing apparatus which performs a sequence of operations to produce a computer-implemented process such that the instructions executing on the computer or other programmable apparatus implement the functions specified by the blocks in the flowcharts and / or block diagrams.

[0151] Furthermore, the teachings herein may be implemented in the form of a computer program product, where the computer program product is stored on a storage medium and includes a plurality of instructions for causing a computing device to perform the methods described in the embodiments of the present disclosure.

[0152] An element or step described in connection with one embodiment may be combined with or substituted by an element or step described in connection with another embodiment. Other variations to the disclosed embodiments can be understood and realized by those skilled in the art in practicing the claimed invention from a study of the drawings, the disclosure and the appended claims. In the claims, the word "comprises" does not exclude other elements or steps, and the singular does not exclude a plurality. A single processor or other unit may fulfill the functions of several items recited in the claims. The mere fact that certain means are recited in mutually different dependent claims does not indicate that a combination of these means cannot be used to advantage. A computer program can be stored or distributed on a suitable medium, such as an optical storage medium or a solid-state medium, provided together with or as part of other hardware, but also in other forms, such as via the Internet or other wired or wireless communication systems. Any reference signs in the claims should not be interpreted as limiting the scope.

Claims

1. Receiving imaging data representing the volume of the anatomical structure of a subject, wherein the received imaging data includes at least one unidentified anatomical feature structure of interest; Using a machine learning model that implements a segmentation technique for identifying at least one anatomical feature structure of interest in the received imaging data, in order to identify the anatomical feature structure of interest in the received imaging data; A computer-implemented method comprising: The machine learning model is trained using a series of training data sets, each training data set including training imaging data representing the volume of the anatomical structure to be trained; The training comprises, for each training data in the series of training data sets: Receiving the training data set and ground truth indication information for identifying the anatomical feature structure of interest in the training data set; Determining at least one control position in the training data set; Generating a map for the training data set by applying a spatial function to at least one control position within the training data set, wherein the spatial function specifies the spatial distribution of at least one training region with respect to at least one control position within the training data set associated with the map, the map imposing a penalty on the learning error in the at least one training region, and the at least one training region includes the at least one unidentified anatomical feature structure of interest in the training data set associated with the map; Training the machine learning model using the training data set among the series of training data sets and the corresponding map of the training data set; And A computer-implemented method.

2. The machine learning model: Uses the segmentation technique to determine where the identified anatomical feature structure of interest is located in the imaging data, and Generates an indicator indicating where the identified anatomical feature structure of interest is located in the imaging data. The computer-implemented method according to claim 1, which identifies the anatomical feature of interest structure by this means.

3. The computer-implemented method according to claim 1 or 2, wherein the at least one control position does not overlap with the anatomical feature of interest structure.

4. The step of using the machine learning model includes the step of using an additional map as an input to the machine learning model for identifying at least one anatomical feature of interest structure in the received imaging data, and the additional map is generated from the received imaging data. The computer-implemented method according to claim 1.

5. The spatial overlap between adjacent training regions specified by the map generated for each training data set causes the machine learning model to penalize learning errors in at least one priority training region more than learning errors in non-overlapping training regions of the training data set and / or other regions of the training data set. The computer-implemented method according to claim 1, which defines at least one priority training region in the training data set for use.

6. The received imaging data corresponds to the basal region of the subject's heart, and the at least one anatomical feature of interest structure to be identified using the trained machine learning model includes at least one anatomical boundary surface between adjacent chambers of the subject's heart. The computer-implemented method according to claim 1.

7. The computer-implemented method according to claim 1, wherein the at least one control position is identified based on the result of an initial segmentation model used to identify the at least one control position.

8. The at least one control position is the centroid of the heart chamber, and / or the endpoints and / or junctions of the ventricular and / or atrial muscle structures that define at least one boundary surface between each chamber of the heart The computer-implemented method according to claim 7, which includes.

9. The spatial distribution of the at least one training region is defined by at least one parameter of the spatial function, and the at least one parameter is based on at least one dimension of at least one previously identified anatomical feature structure in the training data set. The computer-implemented method according to claim 1.

10. The spatial function includes a first Gaussian-based function centered at an origin defined by the at least one control position in the training data set, and the spatial distribution of the at least one training region defined by the first Gaussian-based function is distal from the origin. The computer-implemented method according to claim 1.

11. The computer-implemented method according to claim 10, wherein the first Gaussian-based function includes an inverse Gaussian function.

12. The volume includes at least a part of the heart, and the first Gaussian-based function is centered at the centroid of at least one chamber of the heart. The computer-implemented method according to claim 10 or 11.

13. The spatial function includes a second Gaussian-based function, the second Gaussian-based function specifies a spatial distribution indicating at least one training region associated with itself, and the spatial distribution indicating at least one training region associated with the second Gaussian-based function overlaps adjacent control positions in the training data set. The computer-implemented method according to claim 1.

14. The volume includes at least a part of the heart, and the spatial distribution of the at least one training region defined by the second Gaussian-based function includes a line connecting adjacent endpoints and / or junctions of the ventricular and / or atrial muscle structures defining at least one boundary surface between the chambers of the heart. The computer-implemented method according to claim 13.

15. The loss function used to impose a penalty on the learning error is modified by the map, and the loss function is based on the difference between the measurement value for at least one pixel or voxel of the training imaging data and the ground truth value. The computer-implemented method according to claim 1.

16. including the step of training the machine learning model, The spatial function generates a set of loss values, the set of loss values indicating a spatial distribution of at least one training region and indicating a loss function for imposing a penalty on a learning error in each pixel or voxel of the training data when applied to each pixel or voxel. The computer-implemented method according to claim 1.

17. A non-transitory machine-readable medium storing instructions executable by at least one processor, the instructions causing the at least one processor to receive imaging data representing a volume of a subject's anatomical structure, wherein the received imaging data includes at least one unidentified anatomical feature structure of interest, and In a non-transitory machine-readable medium, use a machine learning model that implements a segmentation technique for identifying at least one anatomical feature structure of interest in the received imaging data to identify the anatomical feature structure of interest in the received imaging data. The machine learning model is trained using a series of training data sets, each training data set including training imaging data representing a volume of the anatomical structure to be trained. The training is for each training data set of the series of training data sets receive the training data set and a ground truth indication for identifying the anatomical feature structure of interest in the training data set. determine at least one control position within the training data set. generate a map for the training data set by applying a spatial function to at least one control position within the training data set, the spatial function specifying a spatial distribution of at least one training region relative to the at least one control position within the training data set associated with the map, the map imposing a penalty on a learning error in the at least one training region, the at least one training region including the at least one unidentified anatomical feature structure of interest in the training data set associated with the map. The machine learning model is trained using a training data set among the series of training data sets and a corresponding map of the training data set. Having this, A non-transitory machine-readable medium.

18. At least one processor communicably coupled to an interface, the interface receiving imaging data representing a volume of a subject's anatomical structure, the received imaging data including at least one unidentified anatomical feature structure of interest, at least one processor; A non-transitory machine-readable medium storing instructions readable and executable by the at least one processor, the instructions causing the at least one processor to use a machine learning model that implements a segmentation technique for identifying at least one anatomical feature structure of interest in the received imaging data to identify anatomical feature structures of interest in the received imaging data. An apparatus having The machine learning model is trained using a series of training data sets, each training data set including training imaging data representing a volume of an anatomical structure to be trained. The training includes, for each training data in the series of training data sets, Receiving the training data set and ground truth indication information for identifying the anatomical feature structure of interest in the training data set; Determining at least one control position in the training data set; Generating a map for the training data set by applying a spatial function to at least one control position in the training data set, the spatial function specifying a spatial distribution of at least one training region with respect to at least one control position in the training data set associated with the map, the map penalizing a learning error in the at least one training region, the at least one training region including the at least one unidentified anatomical feature structure of interest in the training data set associated with the map. A step of training the machine learning model using a training data set among the series of training data sets and a corresponding map of the training data set having an apparatus

Images