Systems and methods for identifying coronary artery disease - Patents.com

Abstract

A method for identifying coronary artery disease is disclosed. The method includes receiving contrast-enhanced cardiac CT data indicative of a contrast-enhanced cardiac CT scan performed on a patient, and identifying a plurality of seed points in the contrast-enhanced cardiac CT data that are expected to correspond to locations in the patient's cardiac arteries by analyzing the contrast-enhanced cardiac CT data using machine learning. The method also includes generating data indicative of transverse image slices of the patient's cardiac arteries using the contrast-enhanced cardiac CT data and the identified seed points, analyzing the transverse image slice data using machine learning to generate artery inner wall data and artery outer wall data indicative of predicted inner and outer walls, respectively, of the patient's coronary arteries, and identifying the presence or absence of coronary artery disease using the predicted inner and / or outer walls of the patient's coronary arteries. A corresponding system is also disclosed.

Description

[Technical field]

[0001] The present invention relates to systems and methods for identifying coronary artery disease, and to systems and methods for characterizing identified coronary artery disease. [Background technology]

[0002] Atherosclerosis is a disease of the coronary arteries in which atheromatous plaque ("plaque") builds up abnormally on the inner lining of the artery wall. Severe buildup of plaque can result in a narrowing of the artery, called arterial stenosis, which can result in reduced blood flow. Arterial stenosis associated with the coronary arteries can lead to heart attacks and death.

[0003] Known systems for identifying coronary artery disease determine a predicted radius from the artery centerline to the artery's inner wall and use this to identify disease, however, this technique has limitations in that it provides limited information about the disease. Summary of the Invention

[0004] According to a first aspect of the present invention there is provided a method for identifying coronary artery disease, comprising the steps of: receiving contrast-enhanced cardiac CT data indicative of a contrast-enhanced cardiac CT scan performed on a patient; analyzing the contrast-enhanced cardiac CT data using machine learning to identify a plurality of seed points in the contrast-enhanced cardiac CT data that are predicted to correspond to locations in a cardiac artery of the patient; generating data indicative of a transverse image slice of a cardiac artery of the patient using the contrast-enhanced cardiac CT data and the identified seed points; generating artery inner wall data and artery outer wall data indicative of predicted inner and outer walls, respectively, of the patient's coronary arteries by analyzing the transverse image slice data using machine learning; identifying the presence or absence of coronary artery disease using the predicted inner and / or outer walls of the patient's coronary arteries; A method is provided, comprising:

[0005] In one embodiment, identifying the plurality of seed points includes identifying a plurality of predicted seed points expected to correspond to locations on the coronary arteries by analyzing the contrast-enhanced cardiac CT data using machine learning, and generating a plurality of candidate seed points by applying a radiodensity test to the predicted seed points. The radiodensity test may include filtering the contrast-enhanced cardiac CT data to pass predicted seed points having associated radiodensity values ​​within a defined parameter range. In one embodiment, the radiodensity test is a Hounsfield Unit test, and the defined parameter range is a Hounsfield Unit value of 100 to 600.

[0006] In one embodiment, the method includes determining a seed point predicted to correspond to a location on the coronary artery by predicting an estimated direction from an instantaneous candidate seed point to another seed point of the coronary artery to identify a line representing the coronary artery. Determining a seed point predicted to correspond to a location on the coronary artery may include selecting a candidate seed point from a plurality of candidate seed points using the predicted estimated direction to another seed point of the coronary artery. Predicting an estimated direction from the instantaneous seed point to another seed point to identify a line representing the coronary artery may be performed using machine learning.

[0007] In one embodiment, the line representing the coronary artery is a line representing the centerline of the coronary artery.

[0008] In one embodiment, the step of predicting an estimated direction from an instantaneous seed point to another seed point includes starting from a seed point at a predicted end of the coronary artery remote from the aorta or a seed point adjacent to the predicted end, and successively predicting seed points from the remote end to the aorta.

[0009] In one embodiment, the method includes using machine learning to detect intersections between the coronary arteries and the aorta by performing an aortic segmentation process on the contrast-enhanced cardiac CT data, for example using machine learning, to predict a location of the ascending aorta in the contrast-enhanced cardiac CT data, and using the predicted location of the ascending aorta and the identified lines representing the coronary arteries. The method may include determining whether a coronary artery joins the aorta based on whether it extends to a location within a specified distance, such as 4 mm, of the ascending aorta.

[0010] In one embodiment, the method includes using a branch detection algorithm to detect branches on a primary coronary artery that were not initially identified as a possible centerline.

[0011] In one embodiment, the method includes a representative arterial line labeler that associates labels identified using machine learning, which may include at least one classifier, with the identified arteries. The at least one classifier may include an adaptive boosting (AdaBoost) algorithm to enhance performance of the classifier.

[0012] In one embodiment, the at least one classifier may be configured to classify the coronary artery based on a plurality of major coronary artery features, which may include a location of an end of the coronary artery remote from the aorta and a directional vector indicating a plurality of different locations along the coronary artery, which may be disposed at a location adjacent to the aorta, a location substantially in the middle of the vessel, and a location at or adjacent to the end of the coronary artery remote from the aorta.

[0013] In one embodiment, the method includes modifying parameter ranges used for the radiation concentration test if a determination is made that the identified coronary arteries are inaccurate or incomplete. If the determination is that an insufficient coronary artery has been identified, the method may include widening the parameter ranges to increase the number of candidate seed points.

[0014] In one embodiment, the transverse image slice is a slice acquired perpendicular to the coronary arteries.

[0015] In one embodiment, generating inner and outer arterial wall data by analyzing the transverse image slice data using machine learning includes training a machine learning component using ground truth data indicative of exemplary transverse image slices including inner and outer arterial walls and imaging artifacts indicative of coronary artery disease.

[0016] In one embodiment, the ground truth data includes a plurality of cross-sectional image data slices for each of a set of a plurality of defined points along the ground truth coronary artery, the image slices including at least one image data slice before the defined point and at least one image data slice after the defined point.

[0017] In one embodiment, the method includes determining a luminal cross-sectional area using the determined inner wall data, and identifying a stenosis using the determined luminal cross-sectional area.

[0018] In one embodiment, the method includes determining, after each arterial bifurcation, a baseline cross-sectional area calculated by fitting a linear regression line to the arterial portion, the linear regression line indicating a linearly decreasing baseline cross-sectional area after the arterial bifurcation, and identifying a stenosis based on a comparison of the determined cross-sectional area to the baseline cross-sectional area according to the linear regression line.

[0019] In one embodiment, the method includes identifying a stenosis when a comparison of the determined cross-sectional area to a reference cross-sectional area indicates a defined proportional difference.

[0020] In one embodiment, the step of identifying the presence or absence of coronary artery disease includes determining a gap area between the determined inner wall data and outer wall data, and analyzing characteristics of the gap area to characterize the coronary artery disease.

[0021] The characteristics of the gap region may include the radiodensity of voxels associated with the gap region.

[0022] In one embodiment, identifying the presence or absence of coronary artery disease includes identifying high-risk plaque, which may include low-attenuating plaque and / or punctate calcification.

[0023] The method may include applying a radiodensity test to identify candidate voxels in the gap predicted to be associated with a calcified plaque, and applying a connected component analyzer to identify punctate calcifications by associating associated voxels together as a calcification volume. The step of identifying punctate calcifications may further include applying a size test to each identified calcification volume, such that a calcification volume is identified as a punctate calcification if it has a diameter less than a predetermined amount, such as 3 mm.

[0024] In one embodiment, the step of identifying the presence or absence of coronary artery disease includes identifying positive remodeling based on whether the radial dimension, cross-sectional area, and / or volume of the gap is greater than a specified amount or a specified percentage compared to a normal coronary artery, such as about 10% greater than a normal vascular gap.

[0025] In one embodiment, the method includes storing different resolution versions of the received contrast-enhanced cardiac CT data and / or facilitating conversion of the received contrast-enhanced cardiac CT data to a lower resolution such that coronary artery disease analysis may be performed on the selected resolution version, where the resolution of the contrast-enhanced cardiac CT data is selected based on desired accuracy and speed of analysis.

[0026] In one embodiment, the method includes enabling a user to edit the arterial inner wall data.

[0027] In one embodiment, the method includes enabling user editing of the arterial wall data, for example by displaying a number of control points representing the arterial wall and enabling the user to move one or more of the control points, thereby facilitating manual modification of the displayed arterial wall by using an interface screen.

[0028] In one embodiment, the method includes enabling a user to edit the arterial inner wall data by facilitating selection of a vascular stenosis, and modifying the arterial inner wall data based on the selected stenosis. Multiple selectable stenosis ranges may be provided.

[0029] In one embodiment, the method includes enabling user editing of the arterial inner wall data by facilitating selection of a vessel segment and / or a level of stenosis at a particular point within the vessel.

[0030] In one embodiment, the method includes enabling user editing of the arterial inner wall data by facilitating selection of the stenosis level of individual lesions.

[0031] In one embodiment, the method includes modifying the arterial inner wall data based on the selected stenosis by performing an iterative process that includes modifying the lumen area of ​​the selected slices and then recalculating the vascular stenosis.

[0032] In one embodiment, if a reduction in the stenosis level is proposed, the step of modifying the luminal area of ​​the selected slices includes reducing the stenosis level of each slice above a maximum level associated with the selected stenosis level by enlarging the luminal area of ​​the slice.

[0033] In one embodiment, if an increase in the stenosis level is proposed, the step of modifying the luminal area of ​​the selected slice includes identifying the slice with the greatest stenosis and increasing the stenosis level of the slice above a minimum level associated with the selected stenosis level by reducing the luminal area of ​​the slice.

[0034] In one embodiment, the iterative process includes three iterations.

[0035] In one embodiment, if the stenosis level of the vessel segment is determined to be 0%, the system may be configured to identify potential stenotic lesions by applying a disease machine learning component to the vessel segment that is trained to recognize disease in coronary vessels.

[0036] According to a second aspect of the present invention, there is provided a system for identifying coronary artery disease, comprising: receiving contrast-enhanced cardiac CT data indicative of a contrast-enhanced cardiac CT scan performed on the patient; a vascular seed detector that uses machine learning to analyze received contrast-enhanced cardiac CT data representative of a contrast-enhanced cardiac CT scan performed on the patient to identify a plurality of seed points in the contrast-enhanced cardiac CT data that are expected to correspond to locations in the patient's cardiac arteries; a vessel wall segmenter that uses the contrast-enhanced cardiac CT data and the identified seed points to generate data indicative of transverse image slices of the patient's cardiac arteries, and analyzes the transverse image slice data using machine learning to generate artery inner wall data and artery outer wall data indicative of predicted inner and outer walls, respectively, of the patient's coronary arteries; a disease assessment unit for identifying the presence or absence of coronary artery disease using the predicted inner and / or outer walls of the patient's coronary arteries; A system is provided, comprising:

[0037] The present invention will now be described, by way of example only, with reference to the accompanying drawings, in which: [Brief description of the drawings]

[0038] [Figure 1] FIG. 1 is a schematic block diagram of a system for identifying (in this example, characterizing) coronary artery disease, according to one embodiment of the present invention. [Diagram 2] FIG. 2 is a schematic block diagram of a coronary artery disease analyzer of the system shown in FIG. 1. [Diagram 3] FIG. 3 is a schematic block diagram of a vascular seed detector of the coronary artery disease analyzer shown in FIG. 2. [Figure 4] 3 is a schematic block diagram of a centerline tracker of the coronary artery disease analyzer shown in FIG. 2. [Diagram 5] FIG. 3 is a schematic block diagram of a vessel wall segmenter of the coronary artery disease analyzer shown in FIG. 2. [Figure 6] 1 is an exemplary cross-sectional depiction of a portion of a cardiac artery imaged using a CTCA scan. [Figure 7] FIG. 3 is a schematic block diagram of a stenosis assessment component of the disease assessment unit of the coronary artery disease analyzer shown in FIG. 2. [Figure 8] FIG. 3 illustrates a high-risk plaque assessment component of the disease assessment unit of the coronary artery disease analyzer shown in FIG. 2. [Figure 9A] FIG. 1 is a flow diagram illustrating a method for identifying (in this example, characterizing) coronary artery disease according to one embodiment of the present invention. [Figure 9B] FIG. 1 is a flow diagram illustrating a method for identifying (in this example, characterizing) coronary artery disease according to one embodiment of the present invention. [Figure 10] FIG. 2 is a depiction of a vessel wall editing screen of a system for identifying (in this example, characterizing) coronary artery disease in accordance with one embodiment of the present invention. [Figure 11] 1 is a depiction of an alternative vessel wall editing screen of a system for identifying (in this example, characterizing) coronary artery disease in accordance with an embodiment of the present invention. [Figure 12] FIG. 12 is a representation of the alternative vessel wall edit screen shown in FIG. 11 after selection of a new stenosis level. [Figure 13] FIG. 13 is a diagrammatic depiction of the iterative process performed by the system after selection of a new stenosis level. [Figure 14] FIG. 2 is a flow diagram describing steps 202 to 214 of a method for correcting a vascular wall. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0039] The present disclosure relates to a system and method for identifying coronary artery disease, which in the described examples also quantifies and characterizes the identified coronary artery disease (CAD). The disclosed system and method uses coronary computed tomography angiography (CCTA) data. In the described examples, the system and method automatically identifies, quantifies, and characterizes coronary artery disease by detecting and tracking coronary artery centerlines, estimating the locations of the inner and outer walls of the coronary arteries, and using the estimated inner and outer walls in conjunction with analyzing the composition and spatial characteristics of the identified gaps between the inner and outer walls to determine the extent and characteristics of any identified disease.

[0040] The described systems and methods can identify early stages of coronary artery disease by detecting the presence and severity of arterial narrowing (referred to as stenosis) and identifying high-risk plaques, including punctate calcifications, low-attenuating plaques, and positive remodeling of the vessel wall. The output of the system can be in the form of a report highlighting key findings and priority risks.

[0041] Referring to the drawings, FIG. 1 is a schematic block diagram of a system 10 for identifying and characterizing coronary artery disease.

[0042] In this example, the system 10 is configured to interact with multiple providers of cardiac computed tomography angiography (CCTA) data, represented in Fig. 1 by CCTA scanning devices 12a, 12b and associated picture archiving and communication systems (PACS) 14a, 14b. Each PACS system 14a, 14b is configured to manage the capture and storage of medical image data generated by the CCTA scanning devices 12a, 12b and the communication of the medical image data to a medical image data server 18, which in this example is located remotely from the CCTA service provider and accessible over a wide area network such as the Internet 16. In this example, the medical image data server 18 is a Digital Imaging and Communications in Medicine (DICOM) server, although it will be appreciated that any suitable device for receiving CCTA image data and managing the storage of the received CCTA image data is contemplated.

[0043] The DICOM server 18 is configured to store the received CCTA image data in a data store 20, which may include one or more databases. In this example, the system 10 also includes a personal health information (PHI) anonymizer 22, which may be a separate component or a component integrated into the DICOM server 18. The PHI anonymizer 22 is configured to encrypt patient-specific metadata (typically including name, date of birth, and unique ID number) in the received CCTA image data before it is stored in the data store 20. In this way, the patient-specific metadata is still associated with the CCTA image data, but is only accessible to authorized personnel, for example by using login and password data.

[0044] System 10 is configured to enable interaction with system 10 by a number of authorized users, for example, by providing interface devices 24. Each interface device 24 may include any suitable computing device, such as a personal computer, laptop computer, tablet computer, or mobile computing device.

[0045] The system 10 also includes a coronary artery disease (CAD) analyzer 26 in communication with the data storage device 20 and configured to analyze the CCTA image data stored in the data storage device 20 to identify, quantify, and characterize coronary artery disease in the CCTA image data and generate a report indicative of the analysis.

[0046] The system 10 may be configured to store different resolution versions of the received CCTA data and / or facilitate conversion of the received CCTA data to a lower resolution so that a CAD analysis can be performed on the selected resolution version. In this manner, the user can modify the CAD analysis depending on the desired accuracy and speed of analysis.

[0047] System 10 may be configured to facilitate access in any suitable manner using interface device 24. For example, system 10 may be configured to allow access to CAD analysis device 26 through a web browser on interface device 24, with all or most processing operations occurring remotely from interface device 24. Alternatively, system 10 may be configured such that at least some processing operations occur at interface device 24, for example, by providing interface device 24 with at least one software application that performs at least some processing operations on CCTA data stored in data storage device 20.

[0048] In one alternative, instead of providing a distributed system in which CCTA data received from the patient is stored remotely at a network-accessible location, one or more components of the system 10 may be co-located with the interface device 24 and / or the CT devices 12a, 12b such that most or all processing operations and / or storage of the CCTA data occur at the same location.

[0049] As shown in FIG. 1, in this example, the data stored in the data storage device 20 may also be accessed directly by the interface device 24, for example to allow a user of the interface device 24 to view the raw CCTA data.

[0050] Using the interface device 24, a user can facilitate analysis and / or review the results of the analysis of the CCTA data stored in the data store 20. During analysis, the CAD analyzer 26 extracts relevant CCTA data from the data store 20 and performs an analysis process on the CCTA data to identify, quantify and characterize coronary artery disease present in the CCTA image data and generate an associated report.

[0051] CAD analyzer 26 is shown in more detail in FIG. 2, which is a schematic block diagram illustrating the functional components of CAD analyzer 26.

[0052] The CAD analyzer 26 primarily relies on segmentation of the inner and outer coronary artery walls, and the information generated is used in the scan to detect and assess disease burden. To accurately segment the vessel walls, it is first necessary to identify a representative line of the vessel, in this example the centerline. This is achieved by identifying multiple seed points on each centerline that correspond to voxels in the CT volume that are likely to be located on the centerline of the coronary artery. To facilitate this process, contrast is injected into the bloodstream to increase the contrast, in this example increasing the Hounsfield Units (HU) value of the coronary artery compared to the surrounding tissue.

[0053] The CAD analyzer 26 identifies the vascular seed points by using a vascular seed detector 28, which in this example uses multi-scale filtering and supervised machine learning to detect seed points from training data. The functional components of the vascular seed detector 28 are shown in FIG.

[0054] The vascular seed detector 28 comprises a vascular seed machine learning component 30 (in this example, a volumetric convolutional neural network (CNN)) that is trained using ground truth data showing a sufficient number of exemplary coronary artery centerlines.

[0055] It is understood that many points in a CT volume have a non-zero probability of being suitable as a vascular seed point. With this in mind, the vascular seed detector 28 includes a candidate seed point determiner 32 that uses the vascular seed machine learning component 30 to identify a set of predicted seed points present in a sample of CCTA data and then selects candidate seed points from the set of predicted seed points on which to base the centerline tracking and thus the prediction of the centerline of the coronary artery. The candidate vascular seed points are determined from the set of seed points based on one or more predefined constraints, such as seed points having a radiodensity value, such as a Hounsfield Unit (HU) value, above a predefined amount, or a predefined number of seed points above a predefined HU threshold, such as a predefined number of seed points having the highest HU value. In one example, candidate vascular seed points having an HU value between 100 and 600 are selected as candidate seed points.

[0056] The vessel seed machine learning component 30 is trained using reference vessel seed points that represent points that experts believe to be on the coronary artery centerlines.

[0057] CAD analyzer 26 tracks coronary artery centerlines using centerline tracker 34. The functional components of centerline tracker 34 are shown in FIG.

[0058] In this example, the centerline tracker 34 includes a centerline tracking machine learning component 36. The centerline tracker 34 considers the determined candidate seed points, and the centerline direction predictor 38 uses the centerline tracking machine learning component 36 to predict the most likely direction from the instantaneous seed point to the next seed point on the coronary artery in three-dimensional space. In this way, vascular centerline seed points that are likely to be on the currently considered coronary artery are identified. In this example, the centerline tracking process starts with a predicted seed point located at the extreme point on the arterial centerline.

[0059] In this example, the centerline tracking machine learning component 36 is configured to use the data cube and the candidate seed points to predict the most likely direction to the next centerline seed point and based thereon select a centerline seed point from the candidate seed points. The seed points thus identified as located on the coronary artery centerlines are connected together using a seed point connector 40 to define a complete coronary artery.

[0060] An important part of centerline tracking is to detect the intersection points between the coronary arteries and the aorta (referred to as coronary ostia). The system and method are configured to detect the aortic intersection points by using an aortic crossing determiner 42 that uses information indicative of the location of the ascending aorta in conjunction with tracking information defined by the determined centerline, and to determine whether the coronary arteries connect to the aorta based on whether they extend to a location within a specified distance, such as 4 mm, from the wall of the aorta. If they do, it is presumed that the path of the coronary artery extends to an intersection point with the aorta and that they are anatomically connected. In this example, the location of the ascending aorta is determined using a cardiac segmenter 35 by performing an aortic segmentation process on the CCTA data, for example using a convolutional neural network (which may be a Unet or Vnet neural network) trained using the ascending aorta ground truth data.

[0061] The centerline tracker 34 is configured to initially detect the four main coronary arteries (left main (LM), left anterior descending (LAD), left circumflex (LCX), and right coronary artery (RCA)). After detecting the main coronary arteries, the artery branch detector 44 then detects branches on the main coronary arteries that were not initially identified as possible centerlines. The main coronary arteries are easier to track than the branches because they carry more blood and therefore more contrast. Smaller branches are more difficult to track because they have a smaller diameter and carry less blood and contrast. The branch detector 44 uses a branch detection algorithm.

[0062] The branch detector 44 estimates the approximate radius of the vessel by looking at the HU values ​​perpendicular to the vessel centerline direction and finding the borders of the coronary arteries based on the HU values ​​since the HU values ​​are significantly smaller outside the vessel wall. If the borders are located on both sides of the centerline, the diameter of the vessel can be measured.

[0063] Branches are detected based on the rate of change of the measured diameter of the vessel along the length of the centerline. For example, if the measured diameter of the vessel increases by more than 10% along the centerline and then returns to its original size, this is marked as a detected branch, noting that coronary vessels naturally decrease in size from proximal to distal locations. At the coronary ostium, the vessel diameter is approximately 4 mm, while at distal locations the vessel diameter typically decreases to less than 1 mm. Thus, the branch detector 44 detects points along the centerline where another coronary artery branches off by examining the rate of change of the estimated diameter.

[0064] As shown in FIG. 2, the CAD analyzer 26 also includes a centerline labeler 46 configured to label the tracked arterial centerlines with semantically significant labels to help the clinician identify the vessels. The identification process involves labeling the coronary centerlines on the identified structured coronary artery tree by using a centerline labeling machine learning component 48, which in this example includes a supervised classifier. The ground truth centerlines annotated with appropriate labels are used to train the classifier. Many different classification processes may be used, and in this example, the classifier includes an adaptive boosting (AdaBoost) algorithm to enhance the performance of the classifier. In this example, the main features used to label each coronary artery in the classification process are the location of the end of the arterial centerline remote from the aorta, and the average direction vector of the centerline at three different centerline locations: a proximal location (adjacent to the aorta), a location in the middle of the vessel, and a distal location of the artery (at or adjacent to the end of the artery remote from the coronary artery).

[0065] 2, the CAD analyzer 26 also includes an anomaly detector 50 configured to improve the reliability of the centerline tracking process by reconfiguring the vascular seed detector 28 if the analysis performed by the centerline tracker 34 is inaccurate or incomplete, e.g., due to the vascular seed detector 28 generating too many or not enough seed points. After the detected aorta is labeled by the centerline labeler 46, if a determination is made that the identified vessel is inaccurate or incomplete (e.g., the initial configuration of the vascular seed detector fails to detect a major coronary artery such as the RCA), a seed parameter modifier 52 of the anomaly detector 50 communicates with the vascular seed detector 28 to reconfigure the parameters of the vascular seed detector 28. In this example, the seed parameter modifier 52 may communicate with the vascular seed detector 28 to reduce the constraints applied by the vascular seed detector 28 so that more candidate vascular seed points are generated, thereby increasing the probability of detecting a vessel in a subsequent iteration. In one example, this can be achieved by broadening the applied Hounsfield Units (HU) test from 100 to 600 to 50 to 600. A similar approach may be used to reduce the number of candidate vessel seed points if too many seed points are being generated.

[0066] It will be appreciated that the anomaly detector 50 functions as a feedback loop to modify parameters of the vascular seed detector 28 until a suitable set of candidate vascular seed points is generated for detecting all coronary arteries or at least a subset of defined coronary arteries, such as all major coronary arteries (LM, LAD, LCX, RCA).

[0067] Coronary arteries may not be detected due to various factors, for example the concentration of contrast given to the patient may be reduced, resulting in too few seed points to pass the applied HU test.

[0068] In the above example, the feedback loop operates automatically. However, it is understood that manual modification of the vessel tracking parameters or the parameters of the vessel seed detector 28, such as the location threshold parameters indicative of what are considered to be valid locations of the coronary arteries, may be performed in response to the observations. For example, if the results show that not all major coronary arteries are successfully identified, the user may modify the valid location parameters and / or the tracking parameters to make the identification of the major coronary arteries more likely to be successful. In this way, the vessel tracking process is more robust because the parameters of the process are changed based on the observations, and the system is more likely to be able to handle CCTA data with different dynamic ranges.

[0069] After all desired coronary arteries have been sufficiently tracked and labeled, the vessel wall segmenter 54 shown in FIG. 2 uses the tracked centerlines to specifically perform the intra- and extra-vascular wall segmentation process by analyzing the CCTA data associated with the coronary arteries.

[0070] The functional components of the vessel wall segmenter 54 are shown in Figure 5. The vessel wall segmenter 54 uses a wall segmentation machine learning component 56 to generate inner and outer wall lumen masks that can be correlated with the presence or absence of calcified and non-calcified plaque and used to identify coronary artery disease taking into account imaging artifacts.

[0071] In this example, the wall segmentation machine learning component 56 is a supervised volumetric convolutional neural network (CNN) trained using ground truth training data representing a sufficient number of exemplary transverse coronary artery image slices (in this example, image slices perpendicular to the arterial centerline) that includes the inner and outer arterial walls and associated imaging artifacts annotated by medical professionals to cover a wide range of examples of disease in various coronary vessels with varying degrees of disease, as well as a variety of typical imaging artifacts indicative of abnormalities such as vessel dilation.

[0072] In this example, to provide a more complete data set, the wall segmentation machine learning component 56 is trained on points along the vascular centerline by providing multiple cross-sectional image data slices, including at least one slice of image data before each defined point of the centerline and at least one slice of image data after each defined point of the centerline. By providing multiple slices to the CNN in this manner, the CNN can incorporate spatial context to more accurately reflect the characteristics of the disease, since the disease appears across multiple cross-sectional slices. However, it is understood that other implementations are possible, such as one image data slice for each defined point of the centerline. Long axis image data may also be used for training purposes to provide the CNN with additional context for the semantic segmentation process. It is also understood that data augmentation can be used to effectively generalize the medial and lateral wall segmentation process from the training data to unknown cases.

[0073] During analysis of the received CCTA data, the vessel wall segmenter 54 samples slices of the CCTA data (in this example, perpendicular to the determined centerline) using a slice analyzer 58. The vessel wall segmenter 54 uses a wall segmentation machine learning component 56 to predict whether each voxel in each acquired image slice is likely to be part of the inner or outer wall of a coronary artery. Voxels identified in the sample slice as being part of the inner or outer wall of a coronary artery are then linked together using a connected component analyzer 60 that uses a technique to identify adjacent voxels that belong to the inner or outer wall of the vessel, thereby generating an inner or outer wall segment.

[0074] Exemplary depictions of a portion of a cardiac artery imaged using a CT scan after wall segmentation are shown in Figure 6. Figure 6a is a "long axis" view 62 of a portion of an artery including coronary artery disease 63, which in this example has been linearly reprojected for ease of reference. Figures 6b and 6c are transverse cross-sections 66, 76 ("short axis" views) of the coronary artery portion shown in Figure 6a at different points along the coronary artery. Figure 6b represents a cross-section along the first cross-plane 68 shown in Figure 6a. Figure 6c represents a cross-section along the second cross-plane 78 shown in Figure 6a.

[0075] In the sample slice shown in FIG. 6b, the vessel wall segmentor 54 identifies the inner wall 70 and outer wall 72, and it can be seen that at the location indicated by the first cross-section 68, the illustrated artery is not affected by disease and the vessel lumen 74 appears unobstructed.

[0076] In the sample slice shown in FIG. 6c, the vessel wall segmentor 54 identifies the inner wall 80 and the outer wall 82, and at the location indicated by the second cross-section 78, it can be seen that the illustrated artery is affected by disease 63, with the vessel lumen 74 appearing to be partially obstructed.

[0077] After completion of coronary artery wall segmentation, it is understood that the system has sufficient data to define the inner and outer wall configuration of the detected coronary arteries, which can be used to determine the presence or absence of disease by analyzing voxels associated with the interstitial regions between the inner and outer walls.

[0078] For this purpose, as shown in Figure 2, the CAD analysis device 26 comprises a disease assessment unit 86. The functional components of the disease assessment unit 86 are illustrated in Figures 7 and 8.

[0079] In this example, CAD analyzer 26 is configured to assess different types of disease, including an arterial configuration called "stenosis," and the presence or absence of high-risk plaque, including calcified, mixed, or non-calcified plaque, by heuristic techniques based on spatial characteristics of gaps in the vessel wall and Hounsfield unit values.

[0080] FIG. 7 illustrates functional components 88 of the stenosis detection and analysis component of the disease assessment unit 86 , and FIG. 8 illustrates functional components 90 of the high-risk plaque detection and analysis component of the disease assessment unit 86 .

[0081] The stenosis functional component 88 includes a lumen area determiner 92 configured to use the inner and outer wall segmentation data to determine the cross-sectional area defined by the inner wall. The vessel lumen cross-sectional area is used to determine and characterize the stenosis state relative to a healthy state.

[0082] It is understood that after the bifurcation of an artery there is usually a relatively narrow artery section that may result in a false positive stenosis. To avoid this, the stenosis functional component 88 includes a post-bifurcation baseline area determiner 96 that uses a calculated baseline area for the area after the bifurcation instead of removing the arterial area immediately after each bifurcation. The baseline area is calculated by fitting a linear regression line to the area along the length of the lumen after the bifurcation using a linear regression analyzer 98. The result of this process is a piecewise linear structure that is discontinuous at the detected bifurcation points. Since coronary arteries naturally reduce in diameter to encase the myocardium (heart), the piecewise linear structure essentially sets an expected or average diminishing area for each healthy (non-stenotic) vessel. A proportional decrease as defined compared to the expected baseline area indicates that the vessel has contracted from its normal health state and a stenosis is present. It is understood that the stenosis assessment component is also capable of determining regarding arteries that have dilated from their normal health state.

[0083] Also, in this example, the stenosis functional component 88 includes a post-processor 100 configured to perform additional processing on the inner and outer wall segment data (e.g., data smoothing operations) to reduce noise due to local variations in lumen area that are not associated with disease.

[0084] In clinical reporting, stenosis is typically assessed as a quantitative percentage, such as 25-49%, etc. The stenosis category determiner 102 is configured to use the stenosis analysis to generate a percentage stenosis assessment and category value for reporting.

[0085] The high-risk plaque functional component 90 is shown in FIG. 8. High-risk plaque (HRP), also referred to as vulnerable plaque, is an early sign of coronary artery disease in a patient. The disease assessment unit 86 uses a heuristic that is a rule-based analysis of arterial wall segments to detect multiple forms of HRP (in this example, low-attenuating plaque, punctate calcification, and positive remodeling). To this end, the high-risk plaque functional component 90 includes a low-attenuating plaque determiner 104, a punctate calcification determiner 106, and a positive remodeling determiner 114.

[0086] Low attenuation plaques are characterized by Hounsfield Unit (HU) values ​​ranging from -30 to 30 Hounsfield Units and can therefore be detected directly by analysis and thresholding of Hounsfield Units.

[0087] Punctate calcifications are defined as relatively small calcifications surrounded by non-calcified or mixed plaque. To detect punctate calcifications, the punctate calcification determiner 106 first determines voxels predicted to be associated with calcified plaque in the determined diseased region between the inner and outer arterial walls by filtering using a predefined radiodensity measure, such as Hounsfield Units (HU) values ​​greater than 350. The associated voxels are then collectively associated as a calcification volume by the connected component analyzer 108. Since punctate calcifications are characterized by a diameter smaller than 3 mm, a size thresholder 110 is used to detect them. An additional heuristic approach is also used by the non-calcified / mixed plaque determiner 112 to determine whether voxels surrounding the identified punctate calcifications have HU values ​​consistent with non-calcified or mixed plaque.

[0088] Positive remodeling is characterized by compensation of the disease between the inner and outer walls by expansion of the outer wall of the vessel. The positive remodeling determiner 114 is configured to detect this using an inner-outer wall gap determiner that determines whether the gap between the inner and outer walls of the artery has increased by more than a specified amount (e.g., more than 10% of the normal vessel gap). The positive remodeling determiner 114 also includes a gap radiodensity analyzer 118, which in this example is configured to determine whether voxels in the gap are consistent with non-calcified plaque, for example by determining the HU value of the voxels in the gap.

[0089] The determinations made by the disease assessment unit 86 form the basis of the reports generated by the report generator 120 .

[0090] Referring to FIG. 9, this is a flow diagram 130 illustrating an exemplary method for identifying (in this example, characterizing) coronary artery disease using the system shown in FIG.

[0091] The method includes extracting CCTA data from the data store 20, identifying predicted vascular seed points using the vascular seed machine learning component 30, and applying a radiodensity filter to the predicted vascular seed points to identify candidate vascular seed points, as shown in steps 132, 134, and 136. Then, using the centerline tracking machine learning component 36 to identify candidate vascular seed points that are predicted to be located on the coronary artery centerline by starting at a location remote from the aorta and predicting a direction to the next seed point, as shown in step 138.

[0092] The cardiac region in the CCTA data is then segmented to determine the predicted location of the ascending aorta, and the aortic segment is used to predict the crossing point between the main coronary arteries and the aorta, as shown in steps 140 and 142. A branch detection algorithm is then applied to detect the coronary artery branches, as shown in step 144.

[0093] As shown in step 146, a centerline labeling machine learning component is used to label the detected coronary arteries.

[0094] If the coronary arteries are not sufficiently detected, as shown in steps 148 and 150, the radiodensity filter parameters used by the vascular seed detector 28 are modified as necessary, and the vascular seed detection and centerline tracking process shown in steps 136-146 is performed again. This process is repeated until the coronary arteries are sufficiently detected.

[0095] The CCTA data is sampled to generate image slice data that each intersects the detected centerline and extends transversely throughout the coronary artery, as shown in steps 152 and 154. Also, a wall segmentation machine learning component 56 is used to generate inner and outer wall lumen data based on the sampled image slices.

[0096] The inner wall data is used to determine the cross-sectional area defined by the inner wall, and the determined area is used to identify and characterize a stenosis, as shown in steps 156 and 158. The determined gap between the inner and outer walls is also used to determine and characterize the presence or absence of high-risk plaque or positive remodeling.

[0097] Analysis of the inner and outer wall data and the gap between the inner and outer walls is used to display a representation of a cross section of the coronary artery, as shown in steps 160 and 162. Additionally, a report can be generated indicating the presence or absence of stenosis and / or high-risk plaque in the CCTA data.

[0098] The system 10 may be configured to allow a user to edit the displayed vessel inner and optional outer walls and associated inner and outer wall data, for example because the clinician believes that the stenosis assessment provided by the system 10 is incorrect. In one configuration, this may be achieved by manually modifying the associated wall(s) using a suitable interface screen.

[0099] For example, as shown in Figure 10, a vessel wall editing screen 170 may be used. Screen 170 includes a long axis cross-section view 172 of a coronary artery and a transverse cross-section view 174 of the coronary artery showing the inner wall 70 and outer wall 72 of the vessel. Using a wall selection button 175, a user may select the inner wall 70 or the outer wall 72, which causes control points 176 to be displayed that represent the shape of the selected wall. A user may change the shape of the selected wall by selecting and moving one or more of the control points 176, for example using a mouse.

[0100] Although the vessel wall editing screen 170 can be used to modify the vessel's inner wall and optionally outer wall 70, 72, this process is relatively tedious as it requires the user to manually move each associated control point 176 until the desired level of stenosis is achieved for the displayed vessel.

[0101] An alternative vessel wall edit screen 178 is shown in Figure 11. According to this configuration, instead of manually correcting the vessel wall by moving control points individually, the alternative vessel wall edit screen 178 includes a stenosis level drop down box 180 that allows the user to select an appropriate stenosis extent 182 for the displayed vessel or for an individual selected lesion. In response to the selection of a new stenosis extent 182, the system 10 modifies the displayed intravascular vessel wall 70 to match the selected stenosis extent 182.

[0102] As shown in Fig. 11, the stenosis level of the displayed blood vessel is determined to be 0%, but the user selects a new stenosis level of 1% to 24% because he believes that 1% to 24% is a more appropriate stenosis level for that blood vessel. In response, as shown in Fig. 12, a new modified inner wall 184 is displayed with a wider gap between it and the outer wall 72, indicating an increased stenosis level.

[0103] Thus, to modify a displayed vessel, the user need only select the appropriate stenosis area 182 of the displayed vessel or selected lesion, and the system will accordingly modify the inner wall of the displayed vessel or lesion to a configuration that satisfies the selected stenosis area.

[0104] The configuration of the displayed vessel inner wall 184 that meets the selected stenosis level may be determined in any suitable manner, for example, by using ground truth training data associated with varying degrees of diseased vessels.

[0105] It will be appreciated that the vessel wall edit screen 178 may be used to modify the stenosis levels assigned to vessel segments or to modify the stenosis levels assigned to individual lesions.

[0106] In an example where it is desirable to modify the stenosis level assigned to a vessel segment, the system 10 may apply the method illustrated in the iterative process diagram 188 shown in FIG. 13 and the flow diagram 200 shown in FIG. 14 which illustrates steps 202-214 of the method for correcting the vessel walls.

[0107] In one example, a user desires to modify the determined stenosis level of a coronary vessel segment from a current level of 25%-49% to a new level of 1%-24% because the user believes this is an appropriate stenosis level for that vessel segment. As shown in step 202, the user uses the stenosis level drop-down box 180 to select the new stenosis range 182.

[0108] If a reduction in the stenosis level is suggested, the method includes reducing the stenosis level of each slice that exceeds the maximum level for the new stenosis category by slightly expanding the slice, as shown in step 204.

[0109] If an increase in the stenosis level is proposed, the method includes identifying the slice with the greatest stenosis, and increasing the stenosis level of the slice by slightly trimming the slice to above the minimum level for the new stenosis category, as shown in step 206.

[0110] In this example, a reduction in the stenosis level is proposed, so that the stenosis level of each slice is reduced above the maximum level of the new stenosis range by slightly expanding the inner vessel wall associated with the slice.

[0111] Vascular slices are graphically represented in FIG. 13, which shows three exemplary slices 190 in the first and two subsequent passes of the iterative process. For each slice, the size of the vessel inner wall is represented by the size of the circle 194, and the level of stenosis in the slice is represented by the percentage value 196 within the circle. As shown, before the modification of the stenosis level, the tenth and eleventh slices are within the proposed new stenosis range, while the twelfth slice is 30% stenotic, which is above the proposed range. As a result, the lumen area is increased slightly to reduce the stenosis level in the twelfth slice, as shown in step 204. The stenosis of the vessel segment is then recalculated taking into account the modified slice lumen area, as shown in step 208.

[0112] The lumen area can be increased or decreased in any suitable manner, for example, the lumen area can be increased by expanding the inner wall using a morphological operator.

[0113] Alternatively, the vessel wall segmenter 54 may include an alternative wall segmentation machine learning component 56 trained to generate a luminal mask indicative of an increased level of stenosis, and a further alternative wall segmentation machine learning component 56 trained to generate a luminal mask indicative of a decreased level of stenosis. In this configuration, the selected alternative or further alternative machine learning component may be used as a default model depending on whether an increased or decreased level of stenosis is desired.

[0114] In yet an alternative arrangement, the lumen area may be modified through the use of image processing techniques such as region growing and / or other segmentation techniques based on the HU value of the vessel.

[0115] As shown in FIG. 13, in the first round of the iterative process, after recalculating the stenosis, the stenosis level of the 12th slice is reduced to 20% due to the increase in the lumen area. However, by increasing the lumen area of ​​the 12th slice, the stenosis level of the adjacent 11th slice is increased to 30% in the stenosis recalculation step 208 because the lumen area of ​​slice 11 is relatively small compared to the 12th slice. In the subsequent iteration (second round), since the stenosis of the 11th slice is above the proposed range, the lumen area of ​​the 11th slice is slightly expanded to reduce the stenosis level, and the stenosis of the vessel segment is recalculated again taking into account the revised slice lumen area, as shown in step 208. This reduces the stenosis level of the 11th slice to 20%. However, by increasing the lumen area of ​​the 11th slice, the stenosis level of the adjacent 10th slice is increased to 30% in the stenosis recalculation step 208. This is because the luminal area of ​​slice 10 is smaller relative to slice 11. Similarly, the stenosis level of slice 12 is also higher because the luminal area of ​​slice 10 is smaller relative to slice 11.

[0116] After three iterations, the stenosis levels of all slices are expected to fall within the selected proposed stenosis range 182 .

[0117] In some situations, the above method does not allow modification of the stenosis extent on a vessel or vessel segment basis. For example, in a situation where one or more lesions of a vessel or segment are in the 90%-100% range, it is not possible to increase the stenosis level of other lesions of the vessel or segment by increasing the stenosis extent of the vessel / segment, since the stenosis extent of the vessel / segment has already reached the maximum level available. In this situation, the user may modify the inner wall of one or more lesions individually.

[0118] In another situation, a clinician may wish to increase the stenosis level of a vessel or vessel segment, but since the vessel / segment has been determined by the system to have 0% stenosis, it is difficult to determine which slice of the vessel / segment to increase.

[0119] In this situation, a disease machine learning component specifically trained to recognize diseases in blood vessels may be used and applied to the vessels / segments determined to be 0% to identify lesions with possible stenosis. After identifying lesions with possible stenosis, the method described above may be used to broaden the stenosis range applicable to the vessel, vessel segment, or individual lesion(s). The disease machine learning component may be any suitable machine learning component, such as a U-Net trained neural network.

[0120] In an alternative configuration for increasing the stenosis level when a vessel / segment is determined to be 0% stenotic, the wall segmentation process described above is discarded and a new wall segmentation process is performed using a machine learning component such as U-Net trained on ground truth vessel slice images and expert annotated stenosis values. After identifying lesions with possible stenosis, the above method may be used to increase the stenosis range applicable to the vessel, vessel segment, or individual lesion(s).

[0121] In the following claims and in the above description of the invention, unless the context requires otherwise, either by express expression or necessary implication, the word "comprise" or variations such as "comprises" or "comprising" are used in the inclusive sense, i.e., to specify the presence of stated features, but do not exclude the presence or addition of other features in various embodiments of the invention.

[0122] It will be understood that where any prior art publication is referred to in this specification, such reference does not constitute an admission that the publication forms part of the general knowledge in the art in Australia or any other country.

[0123] Modifications and variations that may be apparent to those skilled in the art are also deemed to be within the scope of the present invention.

Claims

1. 1. A method for identifying coronary artery disease, comprising: receiving contrast-enhanced cardiac CT data indicative of a contrast-enhanced cardiac CT scan performed on a patient; analyzing the contrast-enhanced cardiac CT data using machine learning to identify a plurality of predicted centerline seed points in the contrast-enhanced cardiac CT data predicted to correspond to locations on centerlines of the patient's cardiac arteries by: analyzing the contrast-enhanced cardiac CT data using machine learning to identify a plurality of predicted centerline seed points and predicting an estimated direction from an instantaneous centerline seed point to another centerline seed point of the coronary artery, and selecting a predicted centerline seed point from the plurality of predicted centerline seed points using the predicted estimated direction to the other centerline seed point of the coronary artery; generating data indicative of a transverse image slice of the cardiac artery of the patient using the contrast-enhanced cardiac CT data and the identified centerline seed points; analyzing the transverse image slice data using machine learning to generate arterial inner wall data and arterial outer wall data indicative of predicted inner and outer walls, respectively, of the coronary artery of the patient; using the predicted inner and / or outer walls of the coronary arteries of the patient to identify the presence or absence of coronary artery disease; A method comprising:

2. The method of claim 1 , wherein the step of identifying a plurality of predicted centerline seed points comprises applying a radiodensity test to pass predicted centerline seed points having associated radiodensity values ​​within a specified parameter range.

3. The method of claim 1 , wherein the step of predicting a probable direction to another centerline seed point is performed using the predicted centerline seed point and a data cube.

4. The method of claim 1 , comprising using machine learning to detect intersections of the coronary arteries and the aorta.

5. The method of claim 4, comprising a step of determining whether a coronary artery is connected to the aorta based on whether it extends to a position within a specified distance from the ascending aorta.

6. 3. The method of claim 2, further comprising modifying the parameter ranges used in the radiation concentration test if a determination is made that the identified coronary arteries are inaccurate or incomplete.

7. 2. The method of claim 1, comprising: determining a luminal cross-sectional area using the determined inner wall data; and identifying a stenosis using the determined luminal cross-sectional area.

8. 8. The method of claim 7, comprising the steps of: determining, after each arterial bifurcation, a baseline cross-sectional area calculated by fitting a linear regression line to the arterial portion, the linear regression line indicating a linearly tapering baseline cross-sectional area after the arterial bifurcation; and identifying a stenosis based on a comparison of the determined cross-sectional area to the baseline cross-sectional area according to the linear regression line.

9. The method of claim 1 , further comprising the step of allowing a user to edit the arterial wall data.

10. 10. The method of claim 9, further comprising: enabling a user to edit the arterial inner wall data by facilitating selection of a vascular stenosis, the system modifying the arterial inner wall data based on the selected stenosis.

11. 11. The method of claim 10, further comprising: modifying the arterial inner wall data based on the selected stenosis by performing an iterative process including modifying a lumen area of ​​a selected slice and then recalculating a vessel stenosis.

12. 1. A system for identifying coronary artery disease, comprising: receiving contrast enhanced cardiac CT data indicative of a contrast enhanced cardiac CT scan performed on the patient; a vascular seed detector that uses machine learning to analyze received contrast-enhanced cardiac CT data indicative of a contrast-enhanced cardiac CT scan performed on a patient to identify a plurality of centerline seed points in the contrast-enhanced cardiac CT data that are predicted to correspond to locations in the patient's cardiac arteries by: analyzing the contrast-enhanced cardiac CT data using machine learning to identify a plurality of predicted centerline seed points; and predicting an estimated direction from an instantaneous centerline seed point to another centerline seed point of the coronary artery, and selecting a predicted centerline seed point from the plurality of predicted centerline seed points using the predicted estimated direction to the other centerline seed point of the coronary artery. a vessel wall segmenter that uses the contrast-enhanced cardiac CT data and the identified centerline seed points to generate data indicative of transverse image slices of the patient's cardiac arteries, and analyzes the transverse image slice data using machine learning to generate artery inner wall data and artery outer wall data indicative of predicted inner and outer walls, respectively, of the patient's coronary arteries; a disease assessment unit for identifying the presence or absence of coronary artery disease using the predicted inner and / or outer walls of the coronary arteries of the patient; A system comprising:

13. The system of claim 12 , wherein the vascular seed detector is configured to apply a radiodensity test to pass predicted centerline seed points having associated radiodensity values ​​within defined parameter ranges.

14. The system of claim 12 , wherein a vessel tracker is configured to use the predicted centerline seed point and the data cube to predict a probable direction to another centerline seed point.

15. 13. The system of claim 12, wherein the vessel tracker is configured to detect intersections of the coronary arteries and the aorta using machine learning.

16. 16. The system of claim 15, wherein the vascular tracker is configured to determine whether a coronary artery connects to an aorta based on whether the vascular tracker extends to a location within a specified distance from the ascending aorta.

17. The system of claim 13, wherein the vascular seed detector is configured to modify the parameter ranges used for the radiation concentration test if a determination is made that the identified coronary artery is inaccurate or incomplete.

18. The system of claim 12 , wherein the disease assessment unit is configured to determine a luminal cross-sectional area using the determined inner wall data and to identify a stenosis using the determined luminal cross-sectional area.

19. The system of claim 12 , wherein the system is configured to allow a user to edit the arterial interior wall data.

20. The system of claim 12, wherein the system is configured to enable user editing of the arterial inner wall data by facilitating selection of a vascular stenosis, and the system modifies the arterial inner wall data based on the selected stenosis.

21. 21. The system of claim 20, wherein the system is configured to modify the arterial inner wall data based on the selected stenosis by performing an iterative process including modifying a lumen area of ​​a selected slice and then recalculating a vessel stenosis.

Images