Systems and methods for analysing arterial calcium

Abstract

Disclosed herein is a system for analysing arterial calcium. The system identifies calcium lesions, within an organ boundary, from Non-Contrast Computed Tomography (NCCT) images. A feature learning model is then used to generate a set of images based on an origin of each calcium lesion in the NCCT images and orthogonal axes through that origin and to learn, for each image in the set, calcium image features, inferring detection or otherwise of a calcium lesion in the respective image. A feature vector comprising the calcium image features is then formed for each image, and the feature vectors are concatenated to facilitate generation of a prediction for each calcium lesion, of an artery in which the respective calcium lesion is located. Finally, a scorer predicts a calcium score based on the predictions generated from each calcium lesion.

Description

SYSTEMS AND METHODS FOR ANALYSING ARTERIAL CALCIUMTechnical Field

[0001] The present invention relates, in general terms, to systems and methods for analysing arterial calcium. More particularly, the present invention relates to systems and methods for automated computational detection and characterization of arterial calcium lesions, reducing biases associated with traditional classification techniques and manual labelling.Background

[0002] Coronary artery disease (CAD) has become a major public health challenge, contributing to a high mortality rate. CAD is also a major cause of morbidity and contributes to a significant economic burden among Asian countries. Risk factors for CAD include high cholesterol, obesity, smoking, physical inactivity, poor diet, high blood pressure, diabetes and aging. Effective interventions to reduce the burden of CAD are urgently needed, including primary prevention strategies at the population level, and secondary prevention strategies at the individual level.

[0003] The latest National Institute for Clinical Excellence (NICE) guidelines support Computed Tomography Coronary Angiography (CTCA) as a first-line investigation for the diagnosis of CAD. This assessment is backed up by the results of the Prospective Multicentre Imaging Study for Evaluation of Chest Pain (PROMISE) and Scottish Computed Tomography of the Heart (SCOT-HEART) studies. These studies showed that CTCA increases the diagnostic accuracy and reduces radiation exposure as compared to functional stress testing. Conventionally, CTCA procedures typically require a significant amount of time from a healthcare specialist to analyse the scans. Manual analysis also tends to lead to inter-observer variation between specialists of up to 20%.

[0004] There is hence a need for an effective tool for an automated analysis of arterial calcium lesions, by means of analysing one of the following : Agatston scores, aortic stenosis severity, and characterisation of aortic plaque.Summary

[0005] Disclosed herein is a system for analysing arterial calcium, comprising : memory; at least one processor, the memory storing instructions that, when executed by the at least one processor, cause the system to: receive, at a receiver system, a plurality of Non-Contrast Computed Tomography (NCCT) images; identify, using a segmentation model, calcium lesions of interest in the NCCT images within an organ boundary; use a feature learning model to: generate a set of images for each calcium lesion based on an origin, corresponding to a location of the respective calcium lesion in the NCCT images, and orthogonal axes through the origin; identify, for each image in the set of images, calcium image features of the respective image and forming a respective feature vector comprising the calcium image features; form a concatenated feature vector from the feature vectors for all images in the set of images; and form a prediction for each calcium lesion, based on the concatenated feature vector, of an artery in which the respective calcium lesion is located; and predict, using a scorer, a calcium score based on the predictions.

[0006] Advantageously, the system provides an effective tool for an automated analysis of arterial calcium lesions, by analysing one of the following : Agatston scores, aortic stenosis severity, and characterisation of aortic plaque.

[0007] Disclosed herein is also a system for training a model for calculating a calcium score, comprising : memory;at least one processor, the memory storing instructions that, when executed by the at least one processor, cause the system to: receive, at a receiver system, a plurality of Non-Contrast Computed Tomography (NCCT) images; identify, using a segmentation model, calcium lesions of interest in the NCCT images within an organ boundary; use a feature learning model to: generate a set of images for each calcium lesion based on an origin, corresponding to a location of the respective calcium lesion in the NCCT images, and orthogonal axes through the origin; learn, for each image in the set of images, calcium image features of the respective image and forming a respective feature vector comprising the calcium image features; form a concatenated feature vector from the feature vectors for all images in the set of images; and form a prediction for each calcium lesion, based on the concatenated feature vector, of an artery in which the respective calcium lesion is located; and predict, using a scorer, a calcium score based on the predictions.

[0008] In some embodiments, identifying calcium lesions of interest in the NCCT images comprises: generating a first segmentation mask for each of the plurality of NCCT images; and generating a second segmentation mask from each first segmentation mask, by size thresholding each calcium lesion.

[0009] In some embodiments, identifying calcium lesions of interest in the NCCT images comprises receiving, at the receiver system, one or more third segmentation masks (each interchangeably referred to as a "further segmentation mask") corresponding to the NCCT images, the one or more third segmentation masks each corresponding to an organ defined by the organ boundary.

[0010] In some embodiments, the NCCT images are of the heart, and the third segmentation mask is a pericardium segmentation mask.[Oil] In some embodiments, identifying calcium lesions of interest in the NCCT images comprises generating a 2-dimensional (2D), binary all calcium mask and a 3- dimensional (3D), all calcium mask, from one of the second segmentation mask and the third segmentation mask.

[0012] In some embodiments, the scorer is configured to compute an Agatston score from the 2D, binary all calcium mask, wherein the calcium score is predicted also based on the computed Agatston score.

[0013] In some embodiments, the computed Agatston score is computed for each artery.

[0014] In some embodiments, the calcium score is predicted based on a number of calcium lesions in the artery.

[0015] In some embodiments, the orthogonal axes comprise two or more of axial, coronal and sagittal axes.

[0016] Disclosed herein is a method for analysing arterial calcium, comprising : receiving, at a receiver system, a plurality of Non-Contrast Computed Tomography (NCCT) images; identifying, using a segmentation model, calcium lesions of interest in the NCCT images within an organ boundary; using a feature learning model to: generate a set of images based on an origin in the NCCT images and orthogonal axes; identify, for each image in the set of images, features of the respective image and forming a respective feature vector comprising the features; form a concatenated feature vector from the feature vectors for all images in the set of images; and form a prediction for each calcium lesion, of an artery in which the respective calcium lesion is located; and predicting, using a scorer, a calcium score based on the predictions.

[0017] Further disclosed herein is a method for training a model for calculating a calcium score, comprising :receiving, at a receiver system, a plurality of Non-Contrast Computed Tomography (NCCT) images; identifying, using a segmentation model, calcium lesions of interest in the NCCT images within an organ boundary; using a feature learning model to: generate a set of images based on an origin in the NCCT images and orthogonal axes; learn, for each image in the set of images, features of the respective image and forming a respective feature vector comprising the features; form a concatenated feature vector from the feature vectors for all images in the set of images; and form a prediction for each calcium lesion, of an artery in which the respective calcium lesion is located; and predicting, using a scorer, a calcium score based on the predictions.

[0018] Advantageously, the present invention automates identification and scoring of arterial calcium lesions.

[0019] In some embodiments, large calcium lesions, such as bone, are excluded based on known characteristics of such lesions - for example, that bones are typically considerably larger than calcium lesions, such that size thresholding will exclude otherwise confounding results due to the presence of bone in CT images.Brief description of the drawings

[0020] Embodiments of the present invention relate to systems and methods for analysing arterial calcium lesions, and will now be described with reference to the following figures, in which:

[0021] Figure 1 illustrates a workflow of the pre-processing stage, according to an embodiment of the present invention.

[0022] Figure 2 illustrates a workflow of the modelling stage, according to an embodiment of the present invention.

[0023] Figure 3 illustrates a post-processing workflow for the workflow, according to an embodiment of the present invention.

[0024] Figure 4 provides a series of Fl-score plots for the voxel level predictions of each of the arterial classification outputs (LAD, RCA, LCX, and others) against the number of epochs, according to an exemplary embodiment of the present invention.

[0025] Figure 5 provides a confusion matrix for the risk of Coronary Artery Calcification (CAC), as predicted by the feature learning model and by manual means, according to an exemplary embodiment of the present invention.Detailed description

[0026] The following detailed description is merely exemplary in nature and is not intended to limit the invention or the application and uses of the invention. Furthermore, there is no intention to be bound by any theory presented in the preceding background of the invention or the following detailed description.

[0027] Disclosed herein are systems and methods for analysing arterial calcium lesions. The output of the analysis may be a patient's risk for developing coronary artery disease (CAD), and this risk may be evaluated based on a combination of one or more factors, including: Agatston scores, stenosis severity and plaque characterisation.

[0028] Knowledge of the density and extent of calcium deposits and lesions in the coronary arteries, as obtained from Non-Contrast Computed Tomography (NCCT) images, may then enable a calculation of the Agatston score. An Agatston score may be assigned to each calcified area, based on the density of calcium deposits. The calcium density may be multiplied by the area of calcium deposits, and the resulting scores for all areas may be summed up. The Agatston score is often used to assess the likelihood of heart disease in individuals with intermediate risk factors.

[0029] Embodiments of the invention include a process, referred to as Al Calcium Score (AICS). AICS is an artificial intelligence (Al) package that uses deep learning algorithms to accurately label the calcium lesions in CT images, which then enablescalculation of the Agatston score. The model was trained using a large Asian NCCT dataset. Manual annotations were generated for training purposes, by expert human annotators. To increase training data, data augmentation techniques were used involving uniform 3D tilting of the organ corresponding to the organ boundary - that organ will typically be the heart. This results in additional training data that reflects biological variations. This is distinct from data augmentation techniques that do not take biology into account, such as rearrangement of image features and the like.

[0030] AICS provides a pipeline to locate calcium lesions for Al model prediction. The AICS process, which can be embodied by an AICS system implementing that process, involves a pre-processing stage (Figure 1) to identify lesions of interest in CT images, followed by an Al processing stage (Figure 2) to categorise the lesions of interest, and a post-processing stage (Figure 3) to produce an output, including a calcium score report, for interpretation by a physician.

[0031] This AICS model can be provide tools for clinicians to analyse a patient's risk factors (Agatston scores, stenosis severity, and plaque characterization) and evaluate their risk of developing CAD.

[0032] Advantageously, embodiments of the present disclosure address a lack of an automated software for the detection of arterial calcium lesions. This has been tested, but is not limited only to, a pan-Asian population. Further advantageously, the systems and methods according to the present disclosure provide a replacement for a need for manual labelling of arterial calcium lesions, reducing inter-observer bias, which may be associated with traditional classification techniques.

[0033] In accordance with an aspect of the disclosure, a system for analysing arterial calcium lesions is provided. The system comprises a memory, configured to store instructions that are executable by at least one processor to train the Al model of AICS and to implement the AICS process to new data for detecting and characterizing calcium lesions. By executing the instructions, the system then will receive a plurality of medical images and identify calcium lesions of interest in the images. A feature learning model then generates a set of images (typically along each axis), learns calcium image features of each image in the set, and generates a respective feature vector for each image - "calcium image features" are image features learned by the feature learning model from a set of training images showing one or more calciumlesions. The calcium image features may be features that distinguish images, or regions of images, showing a calcium plaque from images, or regions of images, that do not show a calcium plaque. The calcium image features may also be specific to particular parts of the vasculature, such as the left anterior descending artery (LAD), the left main artery (LM), the left circumflex artery (LCA / LCX) or the right dominant artery (RDA). The feature vectors (e.g. along each of three axes of a CT scan) are then concatenated into a concatenated feature vector from which a prediction is formed for each calcium lesion. A calcium score is then generated based on the predictions. The set of instructions may be referred to as a training and learning model, and the instructions may be divided into three stages: a pre-processing stage, a modelling stage, and a post-processing stage.Pre-Processing Stage

[0034] Figure 1 illustrates a workflow of the pre-processing stage, according to an aspect of the disclosure. A plurality of NCCT images taken of a patient are received at a receiver system. These images may be transmitted using the Digital Imaging and Communications in Medicine (DICOM) standard, which is an international standard used for storing, exchanging and transmitting medical images and data, for enabling integration of medical imaging devices such as scanners, servers, workstations, printers, and network hardware devices. These images may be stored in a series of 0's and 1's, which may be reconstructed to form the image, using the information from a header and image data sets. The DICOM image files may be converted to a NumPy array, which comprises multi-dimensional arrays and matrices.

[0035] In a first step, a value thresholding may be applied to each of the processed NCCT images. Hounsfield Units provide a quantitative measure of radio density, corresponding to the absorption or an attenuation of radiation within a tissue. While surrounding tissues like air and water have little attenuation of radiation and are displayed as low densities (dark fields) in a computed tomography (CT) scan, calcium may present higher attenuation of radiation as compared to air and water and are displayed as relatively higher densities (bright fields) in the same CT scan. Hence, value thresholding provides a means for identifying calcium lesions.

[0036] The quantitative measure of radio density may also directly correspond to the Agatston score and provides a measure of calcium on a coronary CT calcium scan.The Agatston score is conventionally calculated using a weighted value assigned to the highest density of calcification in a given coronary artery.

[0037] The value threshold may be set at 130 Hounsfield Units (HU). Conventionally, an Agatston score of 1 corresponds to a radio density of about 130 HU to about 199 HU. Hence, the value threshold may be selected from the lower limit of the range of radio density at which calcification may be detected - 130 HU.

[0038] Value thresholding generates a first binary segmentation mask (referred to as mask 1 in Figure 1). The first binary segmentation mask (i.e. mask 1) isolates all calcium lesions in the NCCT image, including but not limited to bones and arterial calcium, which may appear on the NCCT image with a radio density of at least 130 HU.

[0039] In a second step, a size thresholding step may be performed along the axial dimension on mask 1, to generate a new mask (referred to as mask 2 in Figure 1). While mask 1 isolates all calcium lesions in the NCCT image, these calcium lesions may appear in a large range of sizes. Coronary artery calcification may begin as microcalcifications, typically in the ranges of about 0.5 pm to about 15.0 pm. This may grow into larger calcium fragments, such as sheet-like deposits of larger than 3 mm. In contrast, a chest bone or a vertebral column, which may appear in a CT scan of the heart, may typically have an average diameter of about 14 mm to about 16 mm.

[0040] The size thresholding may be set at about 10 mm. Size thresholding generates a second binary segmentation mask (i.e. mask 2) which isolates all calcium lesions below 10 mm. This may remove larger calcium lesions, such as bones, isolating the arterial calcium, including microcalcifications and larger sheet-like deposits.

[0041] In some embodiments, a pericardium segmentation mask (referred to as mask 3 in Figure 1) generated by Al Epicardial Adipose Tissue (EAT) is used. The epicardial adipose tissue refers to the fat deposits on the surface of the myocardium and contained entirely beneath the pericardium, surrounding and in direct contact with the major coronary arteries and their branches. The pericardium segmentation mask may be used to remove the calcium lesions outside of the heart.

[0042] In some embodiments, two different binary calcium masks are generated from the pericardium segmentation mask (i.e. mask 3). The first of the two binary calcium masks is a 2-dimensional (2D), binary all calcium mask. The 2D, binary all calcium mask may be generated using the 8-connectivity, that is, 8-connected components along the axial dimension. The 2D, binary all calcium mask is referred to as mask 4 in Figure 1.

[0043] The second of the two binary calcium masks in a 3-dimensional (3D), binary all calcium mask. The 3D, binary all calcium mask may be generated using the 26- connectivity, that is, 26-connected components for 3D binary images. This is referred to as mask 5 in Figure 1.

[0044] In a final step of the pre-processing stage, the 3D, binary all calcium mask and the original NCCT image are resized to isotropic dimensions. This ensures that both the 3D, binary all calcium mask and the original NCCT image are invariant in all directions.

[0045] By applying the various segmentation masks as described according to various exemplary embodiments above, a set of 3D coordinates for each of a plurality of arterial calcium lesions may be obtained. This may be used as an input for the modelling stage.Modelling Stage

[0046] Figure 2 illustrates a workflow of the modelling stage, according to an aspect of the disclosure. The set of 3D coordinates of each of the plurality of arterial calcium lesions may localise a centre voxel of the resized NCCT image - that "centre voxel" is referred to as an "origin" and may correspond to a centre of area in a 2D implementation, or a centre of volume in a 3D implementation, or some other reference location corresponding to a calcium lesion (i.e., a position that enables the set of images, which are orthogonal, to each pass through the lesion - this bears in mind that a calcium lesion may be shaped such that, in a 2D implementation (the same works in 3D), a mid-point having coordinates corresponding to halfway between the two most extreme points in each dimension may not in fact lie within the lesion itself, but all orthogonal images will still pass through the lesion). A centre voxel and its corresponding 3D coordinates may be used to locate the corresponding axial,coronal and sagittal planes of the 2D CT images. The axial, coronal and sagittal planes are mutually orthogonal (i.e., the surface normal for each plane is orthogonal to the surface normal of the other two, such that the surface normal define orthogonal axes). Using this approach, each calcium lesion gives rise to a set of mutually orthogonal images for analysis by the feature learning model.

[0047] The modelling stage comprises a use of a feature learning model. According to one aspect of the invention, the feature learning model may comprise one or multiple modified residual network convolutional neural networks, a hidden block, and a classifier block, for image classification and object detection. The feature learning model may comprise three modified ResNetl8 convolutional neural networks. Each modified ResNetl8 model may comprise four main blocks:An initial layer comprising a stem block- A plurality of residual blocks- A global average pooling layer- A final output

[0048] Alternatively, other Residual Network architecture or appropriate image processing networks may be used.

[0049] When such a modified ResNetl8 model is used, an initial layer comprising a stem block for initial input processing is used. This may comprise a larger kernel size than would be used for the subsequent residual blocks, 16 filters, a batch normalisation layer, an activation layer, and a max-pooling layer. The larger kernel size may be a single 7x7 convolutional layer with a stride of two, the activation layer may be a rectified linear unit (ReLU) layer, and the max-pooling layer may be used to reduce the spatial dimensions of the feature map and retain the most important features. The stem block may generate 16 feature maps with reduced spatial resolutions, from an input image of 600x600 pixels. These feature maps may be passed into the main part of the network, where residual blocks and deeper layers further refine and learn complex features of the data. The output of the stem block is hence fed into the first residual block.

[0050] Each of the four stages of residual blocks may be constructed with two 3x3 convolutional layers, each with a stride of one, a batch normalisation layer and aReLU layer. At each stage of residual blocks, shortcut connections may be integrated in each block to allow for information to bypass the convolution layers, allowing gradients to flow more easily through the network.

[0051] The last residual block may provide an output toward a global average pooling layer, which computes the average of each feature map across the spatial dimensions of the feature map, generating a single 128-feature vector. The feature vectors from each of the three modified ResNetl8 models, corresponding to the axial, coronal and sagittal planes of the NCCT images, may then be concatenated into a single 384 feature vector, to be provided as an input to a hidden block.

[0052] In some embodiments, the hidden block may comprise two fully connected layers with ReLU activation layers. This aids the model in combining and learning abstract and high-level features from the feature vectors output by the last residual block.

[0053] The classifier block consists of a single fully connected layer, which allows the model to learn high-level representations that are optimised for the classification of calcium lesions. The classification outputs may also be referred to as "model predictions" throughout this specification.

[0054] The classification outputs, or model predictions, may be, but are not limited to, one of the following : the left anterior descending artery (LAD), the left main artery (LM), the left circumflex artery (LCA) or the right dominant artery (RDA). Calcium lesions which do not fall under one of these coronary arteries may be labelled as "others" and excluded in the post-processing stage.Post-Processing Stage

[0055] Figure 3 illustrates a post-processing workflow for the workflow, according to an aspect of the present disclosure. 3D coordinates of each arterial calcium prediction may be used to generate arterial-specific labelled segmentation masks. Within the arterial-specific segmentation masks, foreground pixels may indicate the presence of arterial calcium lesions. An Agatston score and number of calcium lesions may be summed up according to a counting method, for generation of a calcium report.

[0056] In a first embodiment according to the present disclosure, calcium lesions are counted based on individual 2D axial slices. Accordingly, if a calcium lesion spans across two 2D axial slices, this would yield a calcium score of 2.

[0057] In a second embodiment according to the present disclosure, the calcium lesions are counted across a 3D volume. Accordingly, if a calcium lesion spans across two 2D axial slices, this would yield a calcium score of 1.

[0058] The scorer may then sum all calcium scores for all lesions, and output the total as the calcium score. Alternatively, the scorer may output a different score for each artery or other component of the vasculature. The scorer may also augment the sum of the calcium scores, or the sum of calcium scores for each artery (or other feature of the vasculature), with a corresponding Agatston score or Agatston scores. Alternatively, the Agatston score may be calculated using the spans and numbers calculated above, and the maximal density of the calcium lesion calculated in a known manner.

[0059] A calcium score report may be generated by first clustering the model predictions according to their classification outputs and obtaining an Agatston score and number of calcium lesions according to one of the above-mentioned methods, for each arterial label. The segmentation masks may be saved and fed for display on a graphical user interface (GUI). By considering a total number of 2D lesions for each of the classification outputs (or model predictions), a value for the Agatston score may be generated.

[0060] A scorer may calculate the Agatston score by means described above. Moreover, the Agatston score may be computed for each artery or other vascular feature. The scorer may comprise a machine learning model trained on data tuples comprising 2D binary all calcium masks (whether of a full NCCT image or images generated for each calcium lesion), and ground truth Agatston scores corresponding to those masks. In other embodiments, the scorer is trained on ground truth Agatston scores and corresponding feature vectors, concatenated feature vectors, predictions, portions of images corresponding to the predictions or other data. The scorer may then calculate the calcium score also based on the computed Agatston score.

[0061] In accordance with another aspect of the disclosure, a method for analysing arterial calcium lesions is provided. The method comprises the following steps:receiving a plurality of medical images, identifying calcium lesions of interest in the images, training a feature learning model to generate a new set of images, learning features of each image of the set of images, forming a respective feature vector and a concatenated feature vector, and forming a prediction for each calcium lesion, to predicting a calcium score based on the predictions.

[0062] The following examples are now provided to assist in detailing how the model was trained, according to various embodiments of the disclosure as defined by the claims. It includes various specific details which are to be regarded as merely exemplary.Example 1 : Study of Calcium Score for a dataset of 1003 patients

[0063] A NCCT dataset of 1003 patients with coronary arterial disease (CAD), from various hospitals in Singapore, including National Heart Centre Singapore (NHCS), Tan Tock Seng Hospital (TTSH) and National University Hospital (NUH) were compiled. The mean age of the patients was 58 ± 11 years, with 33% being female and 72%, 20%, and 8% of the dataset being comprised of Chinese, Indian / Malay, and other ethnicities, respectively. Additionally, 52%, 19%, 69%, and 31% of the patients had comorbidities of hypertension, diabetes, hyperlipidemia, and obesity, respectively.

[0064] Calcium lesion segmentation masks for each arterial label were prepared. The pericardium regions were demarcated at varying intervals along an axial dimension, with an increasing interval at the apex and basal regions. A 3D pericardium segmentation mask was generated by an automated 3D interpolation, and a value thresholding of more than 130 HU was applied on the pericardium region to identify calcium lesions within the heart. An artery label was assigned manually, to each arterial calcium lesion, and any unlabelled lesions were disregarded.Example 2: Training of a AICS model

[0065] The AICS model was trained using a large Asian NCCT data set. Manual annotations were done by one or more skilled persons.

[0066] A sample of 368 patients (mean age 56 ± 11 years; 40% female) were used to train the model. The sample was divided into three sets: 229 patients for training,11 for validation, and 35 for testing. Of these patients, 36% were shown to have very low risk (CAC = 0), 49% had mildly increased risk (CAC = 1-99), 11% had moderately increased risk (CAC = 100-299), and 5% had moderate to severe risk (CAC > 300). The number of labelled 2D axial calcium lesions were as follows: LAD - 929, LM - 101, LCX - 545 and RCA - 940.

[0067] Data augmentation, a statistical technique used to create new data samples from pre-existing data, to improve model optimisation and generalisability, was applied to the data set for training. Uniformly distributed random rotations of the NCCT images are generated by first performing a single random rotation about a vertical axis, then rotating the north pole to an arbitrary position along the x- and y- axis. The labelled calcium lesions for the training and validation data sets were balanced by means of a random weighted data sampler.

[0068] Figure 4 provides a series of Fl-score plots for the voxel level predictions of each of the arterial classification outputs (LAD, RCA, LCX, and others) against the number of epochs. The Fl-scores provide a quantification of both prediction and recall, an indication on how well a model is making predictions. The Fl-score considers the predictions, true positives, false positives, and false negatives. The diamond markers (located in the inlay graphs bounded by a dashed box in each case) in Figure 4 depict the data points used for training, and the triangle markers in Figure 4 depict the data points used for validation.

[0069] By comparing the predicted values, as predicted by the trained feature learning model, to the clinical ground truth for the calcium score, as annotated and calculated manually, a correlation coefficient was calculated to be r = 0.90 (P < 0.0001). The model was found to detect the location of calcium lesions belonging to any one of the three arteries, with accuracies of 0.95 for calcium lesions found in LAD, 0.90 for calcium lesions found in LCX, and 0.80 for calcium lesions found in RCA. The model was also found to demonstrate diagnostic accuracy, a high Fl-score, and an intra-class correlation coefficient (ICC) of 0.80, 0.79 and 0.92 respectively, when predicting CAC risk categories. CAC risk categories used here may be one of the following : low, mildly increased, moderately increased and moderately to severely increased.

[0070] Figure 5 provides a confusion matrix for the risk of CAC, as predicted by the feature learning model and by manual means. A confusion matrix provides an indication of how well the model is performing, by showing a breakdown of its predictions, and the respective frequencies of correct and incorrect predictions.

[0071] While the embodiments in this disclosure refer to NCCT images, it shall be understood that medical images other than NCCT images, in which calcium lesions are detectable, can be used in place of NCCT images, throughout this description and the claims which follow, without a change in process.

[0072] It will be appreciated that many further modifications and permutations of various aspects of the described embodiments are possible. Accordingly, the described aspects are intended to embrace all such alterations, modifications, and variations that fall within the spirit and the scope of the appended claims.

[0073] Throughout this disclosure, unless the context requires otherwise, the word "comprise", and variations such as "comprises" and "comprising", will be understood to imply the inclusion of a stated integer or step or group of integers or steps, but not the exclusion of any other integer or step or group of integers or steps.

[0074] The reference to any prior art in this disclosure is not, and should not be taken as, an acknowledgement or any form of suggestion that the prior art forms part of the common general knowledge.

Claims

CLAIMS1. A system for analysing arterial calcium, comprising: memory; at least one processor, the memory storing instructions that, when executed by the at least one processor, cause the system to: receive, at a receiver system, a plurality of Non-Contrast Computed Tomography (NCCT) images; identify, using a segmentation model, calcium lesions of interest in the NCCT images within an organ boundary; use a feature learning model to: generate a set of images for each calcium lesion based on an origin corresponding to a location of the respective calcium lesion in the NCCT images, and orthogonal axes through the origin; identify, for each image in the set of images, calcium image features of the respective image and forming a respective feature vector comprising the calcium image features; form a concatenated feature vector from the feature vectors for all images in the set of images; and form a prediction for each calcium lesion, based on the concatenated feature vector, of an artery in which the respective calcium lesion is located; and predict, using a scorer, a calcium score based on the predictions.

2. A system for training a model for calculating a calcium score, comprising: memory; at least one processor, the memory storing instructions that, when executed by the at least one processor, cause the system to: receive, at a receiver system, a plurality of Non-Contrast Computed Tomography (NCCT) images; identify, using a segmentation model, calcium lesions of interest in the NCCT images within an organ boundary; use a feature learning model to: generate a set of images for each calcium lesion based on an origin corresponding to a location of the respective calciumlesion in the NCCT images, and orthogonal axes through the origin; learn, for each image in the set of images, calcium image features of the respective image and forming a respective feature vector comprising the calcium image features; form a concatenated feature vector from the feature vectors for all images in the set of images; and form a prediction for each calcium lesion, based on the concatenated feature vector, of an artery in which the respective calcium lesion is located; and predict, using a scorer, a calcium score based on the predictions.

3. The system according to claim 1 or claim 2, wherein identifying calcium lesions of interest in the NCCT images comprises: generating a first segmentation mask for each of the plurality of NCCT images; and generating a second segmentation mask from each first segmentation mask, by size thresholding each calcium lesion to disregard calcium lesions above a threshold size.

4. The system according to any one of claims 1 to 3, wherein identifying calcium lesions of interest in the NCCT images comprises receiving, at the receiver system, one or more further segmentation masks corresponding to the NCCT images, the one or more further segmentation masks each corresponding to an organ defined by the organ boundary.

5. The system according to claim 4, wherein the NCCT images are of the heart, and the third segmentation mask is a pericardium segmentation mask.

6. The system according to claims 4 or 5, wherein identifying calcium lesions of interest in the NCCT images comprises generating a 2-dimensional (2D), binary all calcium mask and a 3-dimensional (3D), all calcium mask, from one of the second segmentation mask and the third segmentation mask.

7. The system according to claim 6, wherein the scorer is configured to compute an Agatston score from the 2D, binary all calcium mask, wherein the calciumscore is predicted also based on the computed Agatston score, and wherein the computed Agatston score is computed for each artery.

8. The system according to claim 7, wherein the calcium score is predicted based on a number of calcium lesions in the artery.

9. The system according to any one of claims 1 to 8, wherein the orthogonal axes comprise two or more of axial, coronal and sagittal axes.

10. The system according to any one of claims 1 to 3, wherein the system is further configured to perform a data augmentation technique, the data augmentation technique involving a uniform 3D tilting of an organ corresponding to the organ boundary.

11. A method for analysing arterial calcium, comprising: receiving, at a receiver system, a plurality of Non-Contrast Computed Tomography (NCCT) images; identifying, using a segmentation model, calcium lesions of interest in the NCCT images within an organ boundary; using a feature learning model to: generate a set of images for each calcium lesion based on an origin corresponding to a location of the respective calcium lesion in the NCCT images, and orthogonal axes through the origin; identify, for each image in the set of images, calcium image features of the respective image and forming a respective feature vector comprising the calcium image features; form a concatenated feature vector from the feature vectors for all images in the set of images; and form a prediction for each calcium lesion, based on the concatenated feature vector, of an artery in which the respective calcium lesion is located; and predicting, using a scorer, a calcium score based on the predictions.

12. A method for training a model for calculating a calcium score, comprising : receiving, at a receiver system, a plurality of Non-Contrast Computed Tomography (NCCT) images;identifying, using a segmentation model, calcium lesions of interest in the NCCT images within an organ boundary; using a feature learning model to: generate a set of images for each calcium lesion based on an origin, corresponding to a location of the respective calcium lesion in the NCCT images, and orthogonal axes through the origin; learn, for each image in the set of images, calcium image features of the respective image and forming a respective feature vector comprising the calcium image features; form a concatenated feature vector from the feature vectors for all images in the set of images; and form a prediction for each calcium lesion, based on the concatenated feature vector, of an artery in which the respective calcium lesion is located; and predicting, using a scorer, a calcium score based on the predictions.

13. The method according to claim 11 or claim 12, wherein identifying calcium lesions of interest in the NCCT images comprises: generating a first segmentation mask for each of the plurality of NCCT images; and generating a second segmentation mask from each first segmentation mask, by size thresholding each calcium lesion.

14. The method according to any one of claims 11 to 13, wherein identifying calcium lesions of interest in the NCCT images comprises receiving, at the receiver system, one or more further segmentation masks corresponding to the NCCT images, the one or more further segmentation masks each corresponding to an organ defined by the organ boundary.

15. The method according to claim 14, wherein the NCCT images are of the heart, and the third segmentation mask is a pericardium segmentation mask.

16. The method according to claim 14 or claim 15, wherein identifying calcium lesions of interest in the NCCT images comprises generating a 2-dimensional (2D), binary all calcium mask and a 3-dimensional (3D), all calcium mask, from one of the second segmentation mask and the third segmentation mask.

17. The method according to claim 13, further comprising computing, at the scorer, an Agatston score from the 2D, binary all calcium mask, wherein the calcium score is predicted also based on the computed Agatston score, and wherein the computed Agatston score is computed for each artery.

18. The method according to claim 17, wherein the calcium score is predicted based on a number of calcium lesions in the artery.

19. The method according to any one of claim 11 to claim 18, wherein the orthogonal axes comprise two or more of axial, coronal and sagittal axes.

20. The method according to any one of claims 11 to 13, further comprising performing a data augmentation technique, the data augmentation technique involving a uniform 3D tilting of an organ corresponding to the organ boundary.

Images