Systems and methods for estimating healthy lumen diameter and stenosis quantification in coronary arteries

Through machine learning algorithms and robust nuclear regression methods, the problem of estimating healthy lumen diameters in non-focal diseases in coronary tree is solved, and the accuracy and efficiency of stenosis evaluation of coronary tree is improved.

CN113440112BActive Publication Date: 2024-11-08HEARTFLOW INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202110698520.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2016-03-16
Filing Date
2017-03-15
Publication Date
2024-11-08
Estimated Expiration
2037-03-15

AI Technical Summary

Technical Problem

The prior art has difficulty in accurately estimating healthy lumen diameters in non-focal disease areas of coronary tree, especially in diffuse, open and bifurcated lesions, resulting in difficulty in estimating the percentage of stenosis.

Method used

Using machine learning algorithms and robust nuclear regression methods, data from individual healthy vascular sites are used to estimate patients' healthy lumen diameters and provide non-invasive assessments by calculating lumen stenosis scores (LNS), including using global nuclear fitting, scoring Regression quantities such as segment fitting and anisotropic nuclear fitting ensure narrow capture of different lengths and positions.

Benefits of technology

Improves the accuracy of estimation of healthy lumen diameters, better than traditional invasive quantitative coronary angiography, especially in diffuse, open and bifurcated lesions, providing safer and time-consuming dissection Stenosis assessment method.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113440112B_ABST
    Figure CN113440112B_ABST
Patent Text Reader

Abstract

A system and method for predicting a healthy lumen radius and calculating a vascular lumen stenosis score is disclosed. A method for identifying a lumen diameter of a patient's vascular system includes: receiving a data set including one or more lumen segments of a plurality of individual known healthy vascular segments; extracting one or more lumen features of each vascular segment; receiving the lumen segments of the patient's vascular system; determining a location of the patient's vascular system; and determining a healthy lumen diameter of a location of the patient's vascular system using the extracted one or more features of each known healthy vascular segment of the plurality of individuals.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] Related Applications

[0002] This application is a divisional application of an invention patent application with application number 201780022642.3, filed on March 15, 2017, and entitled “System and method for estimating healthy lumen diameter and stenosis quantification in coronary arteries”.

[0003] This application claims priority to U.S. Provisional Application No. 62 / 309,376, filed on March 16, 2016, the entire disclosure of which is hereby incorporated by reference in its entirety. Field of the Invention

[0004] Various embodiments of the present disclosure generally relate to imaging and related methods. More particularly, certain embodiments of the present disclosure relate to systems and methods for predicting healthy lumen radius and calculating vascular lumen stenosis scores. Background Art

[0005] Coronary artery disease (CAD) is one of the leading causes of death. CAD can be characterized as an acute event or a gradual event. An acute event may include a plaque rupture that may require immediate care. A gradual event may include an accumulation of plaque that may lead to progressive anatomical stenosis causing ischemia. One of the most widely used noninvasive clinical indicators for diagnosing patients with symptoms of coronary artery disease is the percent stenosis obtained from coronary computed tomography angiography (cCTA). The estimation of the percent stenosis may involve two steps: (1) measurement of the local diameter and (2) measurement of a reference healthy diameter. To measure the percent stenosis, cCTA may provide information about the extent of anatomical stenosis in different regions of the coronary artery tree. The extent of anatomical stenosis in a region of the coronary artery tree may be a clinical measure used to decide between performing invasive angiography and pressure measurements or delaying invasive measurements. In some embodiments, the degree of anatomical stenosis can be explicitly estimated in the clinic (e.g., 0%, 1-30%, 31-49%, 50-69%, 70-100%) or sent to a core laboratory for analysis. Quantitative computed tomography (QCT) and quantitative coronary angiography (QCA) may include methods in which the percentage of stenosis can be estimated as a number between 0 and 100. QCA may involve an invasive procedure evaluated on angiograms, and QCT evaluated on cCTA may be time consuming and is typically performed in a core laboratory. Therefore, there is a desire to provide a safe and less time consuming method for determining the degree of anatomical stenosis in a region of the coronary artery tree.

[0006] Determining the extent of stenosis requires first identifying a reference vessel diameter, such as the healthy lumen diameter. However, estimation of the reference healthy diameter may be challenging in areas of non-focal disease, such as diffuse, ostial, and bifurcation lesions. There is also a desire to estimate the healthy lumen diameter in non-focal lesions.

[0007] The foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. Summary of the invention

[0008] According to certain aspects of the present disclosure, systems and methods for predicting healthy lumen radius and calculating vascular lumen stenosis scores are disclosed. A method for identifying the lumen diameter of a patient's vascular system includes: receiving a data set including one or more lumen segments of known healthy vascular segments of multiple individuals; extracting one or more lumen features of each vascular segment; receiving the lumen segments of the patient's vascular system; determining a location of the patient's vascular system; and using the extracted one or more features of each known healthy vascular segment of the multiple individuals to determine the healthy lumen diameter of the location of the patient's vascular system.

[0009] According to another embodiment, a system for identifying a lumen diameter of a patient's vascular system includes: a data storage device storing instructions for identifying image acquisition parameters; and a processor configured to: receive a data set including one or more lumen segmentation of known healthy vascular segments of multiple individuals; extract one or more lumen features of each vascular segment; receive the lumen segmentation of the patient's vascular system; determine a location of the patient's vascular system; and determine a healthy lumen diameter of a location of the patient's vascular system using the extracted one or more features of each known healthy vascular segment of the multiple individuals.

[0010] According to yet another embodiment, a non-transitory computer-readable medium for use on a computer system containing computer-executable programming instructions for identifying a lumen diameter of a patient's vascular system is provided. The method includes: receiving a data set including one or more lumen segments of known healthy blood vessel segments of a plurality of individuals; extracting one or more lumen features of each blood vessel segment; receiving the lumen segments of the patient's vascular system; determining a location of the patient's vascular system; and determining a healthy lumen diameter of a location of the patient's vascular system using the extracted one or more features of each known healthy blood vessel segment of the plurality of individuals.

[0011] Additional objects and advantages of the disclosed embodiments will be set forth in part in the following description, and in part will be obvious from the description, or may be learned by practice of the disclosed embodiments. The objects and advantages of the disclosed embodiments will be realized and obtained by means of the elements and combinations particularly pointed out in the appended claims.

[0012] It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosed embodiments, as claimed. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate various exemplary embodiments and, together with the description, serve to explain the principles of the disclosed embodiments.

[0014] Figure 1 is a block diagram of an exemplary system and network for predicting healthy lumen radius and calculating a vessel lumen stenosis score (LNS) according to an exemplary embodiment of the present disclosure.

[0015] Figure 2A is a block diagram of an exemplary method 200 for generating estimates of a patient's healthy lumen diameter and a lumen stenosis score according to an exemplary embodiment of the present disclosure.

[0016] Figure 2A is a block diagram of an exemplary method for generating estimates of a patient's healthy lumen diameter and lumen stenosis score according to an exemplary embodiment of the present disclosure.

[0017] Figure 2B is a block diagram of an exemplary method of assessing a patient's vasculature using an LNS according to an exemplary embodiment of the present disclosure.

[0018] Figure 3A is a block diagram of an exemplary method 300 for a training phase for developing a machine learning algorithm for generating an estimate of healthy lumen diameter (which can be used to calculate a lumen stenosis score) according to an exemplary embodiment of the present disclosure.

[0019] Figure 3B According to an exemplary embodiment of the present disclosure Figure 3A An exemplary vascular tree for the machine learning algorithm.

[0020] Figure 3C is an improvement or further training according to the exemplary embodiments of the present disclosure for verification by Figure 3A A block diagram of an exemplary method of training a machine learning algorithm to generate a luminal stenosis score.

[0021] Figure 4 is the use according to an exemplary embodiment of the present disclosure (e.g., Figure 3A Block diagram of an exemplary method 400 for generating a patient-specific luminal stenosis score using a machine learning algorithm (described herein). DETAILED DESCRIPTION

[0022] Reference will now be made in detail to exemplary embodiments of the present invention, examples of which are illustrated in the accompanying drawings. Whenever possible, the same reference numerals are used throughout the drawings to refer to the same or like parts.

[0023] Although indications of lumen stenosis and percentage stenosis are popular non-invasive clinical indicators for diagnosing patients with arterial disease, current methods involve quantitative coronary angiography (QCA) assessed on computed tomography angiography (CTA) data or quantitative coronary angiography (QCA) assessed on angiograms. QCA methods are invasive, and QCT methods are time consuming and are typically performed in core laboratories or clinics. Therefore, there is a desire to provide a safe and less time consuming method for determining the extent of anatomical stenosis in a region of the coronary artery tree.

[0024] The present disclosure relates to providing indications of lumen stenosis, percentage stenosis, and disease non-invasively for a given lumen segment. Existing efforts often involve determining healthy lumen diameter from finding the lumen diameter of a patient's vessel upstream of a lesion and the lumen diameter of a patient's vessel downstream of a lesion. Such methods can capture focal coronary artery disease, where the lesion or stenotic area can be clearly separated from the lumen of a normal or healthy vessel. However, such methods may fail to reliably detect lesions, where there is no clear indication of healthy (relative to diseased) lumen diameter, for example, in diffuse, ostial, and bifurcated lesions. For diffuse and ostial lesions, for example, an area of ​​disease may span a lengthy portion of the vascular system without presenting obvious stenosis in the lumen geometry. In such cases, it may be difficult to discern where the diseased portion of the vascular system may begin and end or what the healthy lumen diameter may be. For bifurcations, even healthy blood vessels may also show a natural decrease in diameter. As a result, healthy lumen diameters are also difficult to determine for vascular bifurcations. Due to the lack of a clear reference lumen diameter, intra-patient estimation / regression may not be able to estimate a reference healthy diameter in these cases.

[0025] To estimate healthy lumen diameter (and thus an indication of luminal stenosis or disease), the present disclosure includes systems and methods for deriving healthy lumen diameter for a vascular site derived from a source other than the patient's own vasculature. In one embodiment, the present systems and methods may determine the patient's healthy lumen diameter using data from an individual's healthy vascular site rather than the patient's vascular site. Alternatively or in addition, a simulated vascular site not necessarily derived from another individual but from a synthetically generated blood vessel may be used to estimate the patient's healthy lumen diameter.

[0026] The present disclosure relates to systems and methods for providing an estimate of a geometric lumen stenosis score (LNS), such as a ratio of a patient's actual lumen radius to an estimated healthy lumen radius. In a simulation embodiment, the LNS may be based on a ratio of a patient's diseased vessel diameter to an estimated healthy diameter.

[0027] As part of generating LNS, the present disclosure includes systems and methods for calculating healthy lumen diameter. In one embodiment, the healthy lumen diameter can be calculated by robust kernel regression or by using a machine learning algorithm. In one embodiment, the robust kernel regression may include multiple regressors, such as global kernel fitting, segmented fitting, and anisotropic kernel fitting. Different series of regressors may be selected to include different lesion locations (e.g., openings, bifurcations) or lesion lengths (e.g., focal or diffuse). In one embodiment, the system and method may include the selection of one or more regressors to ensure the capture of different lengths and locations of lumen stenosis. The selection of regressors may include several factors, including considerations for minimizing the effects of sharp radius changes at vessel branches.

[0028] In one embodiment, a machine learning algorithm may determine the diameter of a healthy lumen relative to a database of healthy sites from a population of individuals rather than from within-patient estimates or regressions. Population-based estimates have been shown to predict the diameter of a healthy site more accurately than within-patient estimates. Such a method can predict the diameter of a healthy site with a correction factor of 0.95. Compared to anisotropic kernel regression methods, machine learning methods can have superior area under the curve (0.9 versus 0.83) and superior operating point sensitivity / specificity for detecting stenosis (90% / 85% versus 82% / 76%). Such methods can also demonstrate superior performance relative to invasive quantitative coronary angiography, which may be due to superior performance in capturing diffuse, ostial, and bifurcation lesions and emphasizing differences in sites with non-focal stenosis.

[0029] In one embodiment, the present disclosure may include a training phase for training a machine learning algorithm, and a generation phase, in which the machine learning algorithm may be used to determine the healthy lumen diameter of a patient of interest. During the training phase, the machine learning algorithm may learn the relationship between upstream and downstream vasculature of a given site of the vasculature, wherein the site may include a healthy vessel site. In other words, the training of such an algorithm may be performed on healthy vessel sites, for example, from manually annotated healthy and diseased sites. An exemplary machine learning algorithm may use a random forest regressor to estimate the healthy lumen diameter of a site using features of vascular sites upstream and downstream of the site. The machine learning algorithm may include a general framework that may identify areas of lumen stenosis in a (coronary) artery, including focal, diffuse, ostial, and bifurcation disease. In an exemplary embodiment, the (coronary) artery may be divided into sites or trunks, wherein each trunk may be associated with features of its corresponding crown (downstream vasculature), root (upstream vasculature), and / or siblings (other child vessels of its parents, if available). One embodiment may include using a machine learning algorithm performed on these features on a database of trunks from a population of individuals to predict the healthy diameter of the trunk. In one embodiment, the machine learning algorithm may be further validated by testing on trunks from a second population of individuals. Such a machine learning algorithm may provide improvements over state-of-the-art techniques in terms of different lesion characteristics.

[0030] During the generation phase, the machine learning algorithm may determine features of the upstream and downstream vasculature of a given location of the patient's vasculature and map those features to an estimate of a healthy vessel radius. By extension, the present disclosure may further relate to determining a luminal stenosis score for a location of a particular patient's vasculature. For example, the generation phase may include then generating a luminal stenosis score for a particular patient that is mapped to a centerline of the patient's vasculature. In one embodiment, the patient's luminal stenosis score may be validated (e.g., against manual annotation of luminal segments of the patient's vasculature) and / or used to update the machine learning algorithm used to determine the patient's luminal stenosis score.

[0031] Healthy lumen diameter and / or LNS can be used in various ways. In one embodiment, LNS can be used as an input to estimate the blood flow reserve fraction (FFR) or sensitivity (e.g., the difference in FFR caused by uncertainty in lumen segmentation). For example, FFR or sensitivity can be calculated via a machine learning algorithm as described in U.S. Application No. 13 / 895,893, the entire disclosure of which is incorporated herein by reference. Such an algorithm may include multiple features as inputs to the machine learning algorithm, including geometric features (e.g., minimum upstream diameter, downstream diameter, etc.), anatomical features (lumen diameter, position in the vascular system, etc.), hemodynamic features (blood viscosity, blood pressure, blood flow, etc.), etc. LNS can be included as a feature in a machine learning algorithm for calculating FFR or sensitivity.

[0032] Another use of LNS may include using LNS to identify trim plane locations so that locations of disease are not trimmed from the model or image. When generating an anatomical model, the model may be trimmed in portions where certainty is lacking, such as in areas that are not clearly imaged. However, when trimming the model, the practitioner may want to avoid areas that may be relevant to understanding vascular disease, such as sites of stenosis. LNS may help the practitioner determine where the model may or may not be trimmed.

[0033] Yet another use of the healthy lumen diameter or LNS may include using the LNS to estimate the ideal lumen diameter in the terminal vessel. In one embodiment, the estimated ideal / healthy lumen diameter may be used to generate a fractal tree, thereby modeling the vessel structure or vessel morphology across the vasculature discernible from the image data. The resulting fractal tree may then allow calculation of downstream resistance to blood flow, for example, by determining healthy vessel area from the healthy lumen diameter and mapping the healthy vessel area to downstream resistance.

[0034] Another application of the healthy lumen diameter or LNS may include using the LNS to estimate areas of disease in a vessel such that an automated pin may be placed distal to an area having an LNS greater than a cutoff value. In other words, the LNS may be used to indicate to a practitioner an area of ​​luminal stenosis, who may initiate a more in-depth study at that area, for example, by initiating a simulated determination of fractional flow reserve (FFR) using the methods described in U.S. Pat. No. 8,315,812, issued to Charles A. Taylor on Nov. 20, 2012, the entire disclosure of which is incorporated herein by reference. In such a use case, the cutoff value may be based on clinician feedback or input such that enough information about luminal stenosis is captured, but the display does not include too many pins.

[0035] By extension, LNS can be used to estimate the location of disease in the patient's vasculature. For example, the lumen radius from the opening to the downstream vasculature can be extracted, and a robust kernel regression with radial basis functions can be used to estimate the healthy lumen radius. The robust kernel regression method can also include modifications to account for natural discontinuities in the lumen radius in bifurcations that may not be detected by an anisotropic kernel. Exemplary modifications to the robust kernel regression can include an anisotropic kernel centered on the bifurcation, where the anisotropic kernel can be convolved with a Gaussian kernel. Such modifications can provide a robust kernel regression that can more reliably estimate the presence of bifurcated lesions.

[0036] Although certain embodiments of the present disclosure are described for purposes of example, with respect to the diagnosis and treatment of coronary artery disease, the systems and methods described herein may be applied to the prediction of an optimal set of image acquisition parameters related to any field of medical imaging.

[0037] Now referring to the accompanying drawings, Figure 1 A block diagram depicting an exemplary system and network for predicting healthy lumen radius and calculating a vessel lumen stenosis score (LNS) according to an exemplary embodiment. In particular, Figure 1 A plurality of physicians 102 and third-party providers 104 are depicted, any of which may be connected to the electronic network 100, including the Internet, via one or more computers, servers, and / or handheld mobile devices. The physicians 102 and / or third-party providers 104 may create or otherwise obtain images of the anatomy of one or more patients. The physicians 102 and / or third-party providers 104 may also obtain any combination of patient-specific information (including age, medical history, blood pressure, blood viscosity, etc.). The physicians 102 and / or third-party providers 104 may transmit the anatomical images and / or patient-specific information to the server system 106 via the electronic network 100. The server system 106 may include a storage device for storing images and data received from the physicians 102 and / or third-party providers 104. The server system 106 may also include a processing device for processing the images and data stored in the storage device.

[0038] Figure 2A General embodiments relate to methods of generating an estimate of healthy lumen diameter or a lumen stenosis score. One way to determine the healthy lumen diameter may include a machine learning approach. Figures 3A-3C Describe the machine learning method for training such a machine, and Figure 4 The application of machine learning methods to estimate the healthy lumen diameter or radius for a specific patient is described.

[0039] Figure 2A is a block diagram of an exemplary method 200 for generating estimates of a patient's healthy lumen diameter and a lumen stenosis score according to an exemplary embodiment. Figure 2AThe method may be performed by the server system 106 based on information, images, and data received from the physician 102 and / or third party provider 104 via the electronic network 100 .

[0040] In one embodiment, step 201 may include receiving input to an algorithm for generating a healthy lumen diameter or LNS. For example, step 201 may include receiving anatomical input, including a surface mesh or a centerline including at least a portion of a patient's vascular system. The surface mesh may include a three-dimensional surface mesh. The centerline may include a centerline located at the center. In one embodiment, the input may include extracting vascular features from lumen segments, such as extracting a coronary artery centerline tree. For example, automatic measurements may be extracted from lumen segments. Any type of lumen segmentation may be used. In one embodiment, a trained CT reader may assess lumen segmentation and may make corrections. In another embodiment, the input may be a vector of radii along with a corresponding bifurcation indicator (0 or 1, depending on whether this point corresponds to a bifurcation) and a parent index for each entry in the vector.

[0041] In one embodiment, step 203 may include selecting an algorithm for determining healthy lumen diameter or LNS. The algorithm may include a kernel regression algorithm or a machine learning algorithm. In one embodiment, the kernel regression algorithm may be represented as a default algorithm. Alternatively, the machine learning algorithm may be set as the default algorithm for determining LNS. In another embodiment, method 200 may represent regression and machine learning algorithms for selection without default settings. The selection may be based on input received from a user, default settings in a particular use case or clinic, and / or patient characteristics.

[0042] In choosing whether to use a kernel regression algorithm or a machine learning algorithm, considerations may include whether the practitioner is evaluating focal or nonfocal lesions. Regression may be adequate for detecting focal lesions, whereas machine learning may be superior to regression in identifying areas of nonfocal stenotic morphology, such as disease in long diffuse lesions, ostial lesions, or lesions that may be present along the entire site.

[0043] In one embodiment, step 205a may include operating a regression algorithm. In one embodiment, the regression algorithm may include a kernel regression algorithm across the lumen region of the path from the opening to the downstream or terminal blood vessel. The stenotic region may be characterized by detecting changes in the vessel radius while progressing from the upstream blood vessel site to the downstream blood vessel site, wherein the radius decreases and then increases along the length of the blood vessel. The diseased region in the vascular system may require a sharp and sudden (violent) or long (diffusion) change in the lumen radius, and the radius may naturally have a sharp decrease (e.g., prescribed by Murray's law) at the bifurcation. In order to explain the various expressions of the diseased region and the stenotic region, the current system and method may include a series of global regressors. The global regressor can be used on the entire vascular system, and the local regressor can be used to analyze the local part of the vascular system of the size of the change. The regression of the range of change can be used to infer multiple indicators, and the LNS can be composed of indicators in a way that will provide a conservative estimate of the LNS to the clinician.

[0044] In one embodiment, step 205a may include operating a kernel regression algorithm with radial basis functions to estimate the healthy lumen radius. Alternatively or in addition, step 205a may include operating a regression algorithm that may include anisotropic kernel fitting, wherein an anisotropic kernel centered at the bifurcation may be convolved with a Gaussian kernel. Anisotropic kernel fitting may account for the natural discontinuity of lumen radius at the bifurcation and more reliably estimate the presence of bifurcation lesions than the kernel regression algorithm described above.

[0045] In one embodiment, step 205a may involve the use of multiple regressors. For example, one exemplary embodiment of step 205a may use three different regressors: a global kernel fit, a piecewise fit, and an anisotropic kernel fit. An exemplary global kernel fit may be defined for each path from the root of the vessel (e.g., the opening) to the leaf, where the healthy radius may be given by:

[0046]

[0047] An exemplary piecewise fit may be defined for each segment between branches, where the healthy radius may be given by:

[0048]

[0049] An exemplary anisotropic kernel fit may be defined for each path from the vessel root to the leaf, but weighted using a sigmoidal function centered on the closest opening designed to minimize the effect of sharp radius changes at branches, e.g.:

[0050]

[0051] The sigmoid function S can be given by the following formula:

[0052]

[0053] as well as

[0054] d o (x',x)=d(x',x o )-d(x,x o )-d(x,x up )

[0055] Once the global health radius is calculated The corresponding indicators can be calculated:

[0056]

[0057] And the lumen stenosis score (λ) can be defined as:

[0058] λ(x)=1–κ(x), if κ<=1,

[0059] Otherwise λ(x)=0.

[0060] In one embodiment, σ may be selected for each of the 15 total regressors. x , σ max and r x,max Exemplary parameter values ​​for each regressor may include the following values, where n may range from 1 to 15.

[0061] σ x =5.0*(1+(n-3)*0.4)

[0062] σ max =200.0*(1+(n-3)*0.4)

[0063] σ r =0.25*(1+(n-3)*0.4)

[0064] k=0.1+n*0.3

[0065] Different regressors can provide different lengths and locations of luminal stenosis.

[0066] In one embodiment, step 205b may include selecting and operating a machine learning algorithm to determine a healthy lumen diameter or LNS. In one embodiment, the machine learning algorithm may use information of vasculature other than the patient's vasculature to determine the healthy diameter. Figure 3A The machine learning algorithm is described in further detail.

[0067] In one embodiment, step 207 may include calculating a lumen stenosis score. The lumen stenosis score may be calculated from the ratio of the actual radius to the healthy radius, e.g.

[0068]

[0069] where r healthy (x) may comprise a theoretical healthy radius of the lumen (e.g., provided by a kernel regression algorithm or a machine learning algorithm), and r(x) may comprise the radius of a maximum inscribed sphere within the lumen. The maximum inscribed sphere within the lumen may be determined by finding the closest point from the centerline of the vessel to the surface of the mesh.

[0070] In one embodiment, step 209 may include validating the calculated luminal stenosis score. For example, the LNS may be validated via direct clinical data outputting luminal stenosis at spatial resolution of the centerline, or an automated vessel labeling tool that may provide a comparison of overall luminal stenosis in the main vessels (pLAD, dLAD), etc. The luminal stenosis score calculated from the machine learning algorithm may be validated against a score calculated from manual annotations.

[0071] There may be several processes for manually annotating diseased segments. For example, a trained reader of cCTA may assess lumen segmentation of a group of patients and identify locations of lumen stenosis (e.g., percent stenosis >= 50%). This process may mimic the process of reading percent stenosis from a CT scan in a clinic, such as visually estimated stenosis rather than assessing a reference diameter and evaluating the ratio of the minimum lumen diameter to the reference diameter. One way to provide confidence in the readings may include a scenario in which each patient's vascular system is assessed by three readers, where only sites with consistent readings may be used for training and testing. For convenience, the coronary artery tree may be divided into sites, where each site may be labeled "disease-prone" or "healthy." The location of the bifurcation may be used as a separator to segment the site. Because manual annotation of diseased sites may be performed on lumen segmentation rather than cCTA, performance may not depend on the algorithm used for centerline detection and lumen segmentation. Various centerline detection and lumen segmentation algorithms may be used, while manual annotation is used to validate the lumen stenosis score calculated from the machine learning algorithm. Step 211 may include updating the kernel regression algorithm or the machine learning algorithm based on the results of the validation.

[0072] Figure 2B is a block diagram of an exemplary method 230 of assessing a patient's vasculature using an LNS according to an exemplary embodiment. Figure 2BThe method may be performed by the server system 106 based on information, images, and data received from the physician 102 and / or third party provider 104 via the electronic network 100. Although the embodiments of the method 230 describe the exemplary use of the LNS, all steps of the method 220 may be performed using a healthy lumen diameter or a healthy lumen radius instead of using the LNS.

[0073] In an exemplary embodiment, step 231 may include receiving an LNS. Steps 233-237 may include using the LNS to determine a resistance of a terminal vessel. For example, step 233 may include estimating an ideal lumen diameter in a terminal vessel of the patient. Step 235 may include generating a fractal tree from the estimate of the ideal lumen diameter, and exemplary step 237 may include determining a downstream resistance to blood flow in the terminal vessel of the patient based on the fractal tree. The resistance may be used to simulate blood flow through the terminal vessel (e.g., step 239).

[0074] In one embodiment, steps 241 and 243 may include exemplary uses for estimating and displaying regions of disease. For example, step 241 may include receiving a threshold LNS, such as a threshold that may indicate the location of disease. Step 243 may include generating a display of a visual indication in a region of a vessel distal to a region having an LNS that exceeds the threshold LNS. As previously discussed, the threshold LNS may be prescribed by a clinician or determined based on clinician feedback to indicate to the physician a sufficient amount of disease region to enable analysis, but not so much disease region that the analysis is difficult.

[0075] Figures 3A-3C Can describe training machine learning methods. For example, Figure 3A Methods for training a machine learning algorithm to perform data-driven estimation of healthy lumen geometry, including features that define a healthy lumen diameter or radius, may be included. Figure 3B A diagram may be included of how a machine learning algorithm may analyze portions of the vasculature to learn characteristics of healthy blood vessels. Figure 3C May include methods for validating trained machine learning algorithms. Figure 4 This may include calculating an estimate of the healthy lumen diameter or radius for a particular patient.

[0076] Figure 3A3 is a block diagram of an exemplary method 300 for developing a training phase of a machine learning algorithm for generating an estimate of a healthy lumen diameter (which can be used to calculate a lumen stenosis score) according to an exemplary embodiment. The method 300 may include training the machine learning algorithm on a database of healthy sites (e.g., a population of healthy vascular trunks obtained from a population of individuals) such that data for a particular patient may be mapped to a population-based healthy lumen diameter. The vasculature may include epicardial vasculature. The method 300 may also include testing or validating the machine learning algorithm using a test vascular site from a second population of individuals (e.g., in Figure 3C Although exemplary method 300 describes an embodiment in which the training data set includes healthy vascular trunks from an imaged anatomy of an individual, alternative methods may use synthetic vascular trunks. Figure 3A The method may be performed by the server system 106 based on information, images, and data received from the surgeon 102 and / or third party provider 104 via the electronic network 100 .

[0077] The method 300 may also include exemplary evaluation of the LNS and any indicators that may be used for validation of the estimate of the healthy lumen geometry. In addition, the LNS may provide an indication of the area of ​​disease, for example, by dividing the local diameter with the estimated healthy diameter and comparing the quotient to a diagnostic threshold of 50%. In one embodiment, a recommendation for treatment may be provided based on the LNS. As an example, if the LNS is less than the cutoff value, sports medicine therapy or exercise may be recommended, while if the LNS is not less than the cutoff value, additional invasive tests / procedures may be recommended.

[0078] In one embodiment, step 301 may include receiving lumen segments of healthy vessel radii. For example, step 301 may include receiving annotated lumen segments. The annotations may be provided by a trained reader (e.g., a trained reader of cCTA or CT scans) that can evaluate the lumen segments for each individual in a plurality of individuals and identify locations of lumen stenosis (e.g., percent stenosis >-50%). Each lumen segment for each individual may be evaluated by multiple readers, and the locations used for training and testing may include locations agreed upon by multiple readers. Optionally, annotations may be performed on cCTA data rather than lumen segments. In such an embodiment, a centerline detection or lumen segmentation algorithm may influence the training of a machine learning algorithm for determining healthy lumen diameter (and LNS).

[0079] In one embodiment, step 303 may include dividing each lumen segment into a trunk-cap-root unit (e.g., Figure 3B). The trunk may include the site of interest, whose healthy diameter may be estimated. The crown may include vasculature downstream of the site of interest, and the root may include vasculature upstream of the site of interest. The machine learning method may also include identifying sibling vessels, which may include child vessels of a parent vessel other than the vessel in which the site of interest is located.

[0080] In one embodiment, step 305 may include defining and / or extracting features of vessel segments, such as where each segment may represent a coronary artery segmentation between bifurcations. For each trunk in a given vascular system, step 303 may include extracting one or more of the following features of the corresponding crown, root, and sibling vessels (when available), such as average, maximum, and minimum lumen area (A), volume (V), length (L), V / A, and V / L. In one embodiment, features of a machine learning algorithm may include estimating local diameter using the largest inscribed sphere. Optional or additional methods may include estimating planar areas, such as the area of ​​the lumen along the normal to the centerline. Some features may not be available for some trunks (e.g., the opening may not have a root unit, and the end may not have a crown unit). In a machine learning approach, such features may be assigned a default special value of -1. In one embodiment, features may be defined for a given position, where each position may represent a coronary artery segmentation between splits. Because the flow rate in a given region may be constant (or the region may be defined such that the flow rate is constant within the region), exemplary step 305 may assume that a healthy vessel may maintain its radius within the region to maintain a homeostatic state of wall shear stress.

[0081] In one embodiment, step 307 may include comparing features from other portions of the vascular tree to the trunk under consideration. For example, a machine learning algorithm may analyze one trunk at a time and use features from the rest of the vascular tree to infer a healthy lumen diameter at the trunk under consideration.

[0082] In one embodiment, step 309 may include inferring the healthy lumen diameter of the vessel segment. For example, the machine learning algorithm may include using random forest regression to predict the healthy lumen diameter. Random forests may be effective and robust for high dimensional and heterogeneous features. Random forests may use an ensemble of decision trees, where each decision tree may be composed of a random subset of features and training data. Each decision tree may map an input feature vector to a continuous variable. The values ​​from the decision trees may be pooled together and averaged to calculate the healthy lumen diameter (d p ) is the final prediction value of the local lumen diameter (d l ) and the predicted healthy lumen diameter, the LNS is estimated as α = (1-dp / d l ) x 100%. An exemplary scenario may include each tree having an average of 5 features. A 5-fold cross validation may be used to evaluate the performance of the selected parameters.

[0083] In one embodiment, subsequent steps may include evaluating or validating the machine learning algorithm. For example, the machine learning algorithm may be updated based on the evaluation / validation. For example, evaluating / validating the machine learning algorithm may include evaluating a random forest regressor (e.g., the random forest regressor of step 309) against manual annotations. In one case, the results of the random forest regressor on a set of trunks (e.g., trunks from various patients) may be estimated by evaluating the sensitivity, specificity, and area under the receiver-operator characteristic (ROC) curve. For validation, sites marked as "disease" by the reader may be considered positive, and if the random forest predicts a stenosis percentage >= 50%, such sites may be further classified as "true positives", or otherwise "false negatives". Similarly, sites that may be marked as healthy may be classified as "true negatives" if the random forest predicts a stenosis percentage <$50%, and otherwise "false positives". Sensitivity (S e ) and specificity (S p ) can be defined as:

[0084] S e =TP / (TP+FN)

[0085] S p =TN / (TN+FP)

[0086] ROC curves can be constructed by evaluating sensitivity and specificity for different cutoff values ​​used to define the site of disease, e.g.

[0087] Figure 3B According to an exemplary embodiment Figure 3A As mentioned above, the vascular tree can be divided into many trunk-crown-root units. Figure 3BAn exemplary coronary artery tree 330 may be included. In one embodiment, a trunk 331 may be defined based on branching points as separators, with the corresponding crowns and roots being the downstream and upstream vasculature, respectively. For the exemplary coronary artery tree 330, epicardial volumes, lengths, diameters, and different ratios may be calculated in the crowns 333, roots 335, and sibling vessels 337 (if available), and various ratios may be assigned as features of a given trunk 331. In one embodiment, a patient's coronary vasculature may be divided into various trunk-crown units, where the trunk may consist of the locations of the coronary arteries and the crowns may consist of the downstream vasculature, where a power law may relate the crown volume and crown length to the trunk area. This embodiment may indicate the presence or absence of luminal stenosis without specifying where disease, such as diffuse disease, exists.

[0088] Figure 3C is an improvement or further training according to the exemplary embodiment for verification by Figure 3A Block diagram of an exemplary method 350 of a machine learning algorithm trained to generate a lumen stenosis score. The validated results from the validation process can provide a basis for the assessment of lesions with complex morphology in addition to providing an accurate estimate of the percent stenosis of the vessel lumen from lumen segmentation. Figure 3C The method may be performed by the server system 106 based on information, images, and data received from the physician 102 and / or third party provider 104 via the electronic network 100 .

[0089] In one embodiment, step 351 may include identifying the Figure 3A ) of individuals (and their vascular locations). In one embodiment, step 351 may include identifying individuals who undergo coronary angiography, where QCA may be used to identify and quantify corresponding diseased locations, such as by an independent expert in a core laboratory. For example, coronary QCA data from a subset of the DeFACTO clinical trial (Clinicaltrials.gov\#NCT01233518) may be used as reference real valid data for the exemplary validation process of method 350 to validate the results of exemplary method 300.

[0090] In one embodiment, step 353 may include determining the healthy diameter and / or presence of disease for the identified vasculature sites of the batch of individuals. For example, a set of geometry-based features including downstream vasculature, upstream vasculature, and sibling vessels may be used to estimate healthy vessel size for a given site. Step 353 may include repeatedly dividing vessel geometry into various "trunk-crown-root" units and using metrics such as epicardial vascular volume and luminal area, which may be known indicators of healthy vessel diameter.

[0091] In one embodiment, step 355 may include calculating a correlation coefficient between the predicted healthy lumen diameter and the corresponding measured healthy lumen diameter. The validation process may also include calculating the mean absolute error and the root mean square error between the predicted healthy lumen diameter and the corresponding measured healthy lumen diameter. The operating point sensitivity and specificity for detecting percent stenosis using the present method may be compared to another method, such as a global kernel regression method or an anisotropic kernel regression method. The validation process may also include calculating and / or comparing a receiver operator characteristic (ROC) curve of the present invention to other methods (e.g., anisotropic kernel regression and / or global kernel regression).

[0092] Figure 4 is a method according to an exemplary embodiment using a machine learning algorithm (e.g., Figure 3A Block diagram of an exemplary method 400 for generating a lumen stenosis score for a particular patient using a machine learning algorithm (e.g., a machine learning algorithm trained on a database of healthy vascular sites from a population of individuals other than the patient). The method 400 may use machine learning to map indices derived from the vasculature of a particular patient to a healthy lumen diameter using a machine learning method, such as a machine learning algorithm trained on a database of healthy vascular sites from a population of individuals other than the patient. While estimating / regressing the healthy lumen diameter from the patient's own vasculature may be useful for detection of focal lesions, the method 400 is more effective for estimating the healthy lumen diameter of non-focal stenoses (e.g., diffuse, ostial, and bifurcation lesions). Figure 4 The method may be performed by the server system 106 based on information, images, and data received from the physician 102 and / or third party provider 104 via the electronic network 100 .

[0093] In one embodiment, step 401 may include receiving a lumen segment of a patient's vasculature. The lumen segment may include a vessel centerline and / or surface mesh representation of the patient's vasculature.

[0094] In one embodiment, step 403 may include dividing the lumen segment into trunk-cap-root units (e.g., Figure 3B ), where the trunk may include a site of interest whose healthy diameter may be estimated, the crown may include vasculature downstream of the site of interest, and the root may include vasculature upstream of the site of interest. Step 403 may also include identifying a sibling vessel of the vessel where the site of interest is located. Exemplary sibling vessels may include child vessels of a parent vessel other than the vessel where the site of interest is located.

[0095] In one embodiment, step 405 may include defining and / or extracting features of the vessel segments. For each trunk, step 405 may include extracting one or more of the following features of the corresponding crown, root, and sibling vessels (when available), such as average, maximum, and minimum lumen area (A), volume (V), length (L), V / A, and V / L. Step 405 may include using the largest inscribed sphere to estimate the local diameter and / or estimate the planar area, such as the area of ​​the lumen along the normal to the centerline. In one embodiment, some features that are not available for some trunks (e.g., the opening may not have root units, and the terminal may not have crown units) may be assigned a default special value of -1. In one example, each estimated segment may represent a coronary artery segment between bifurcations.

[0096] In one embodiment, step 407 may include comparing features from other parts of the vascular tree to the trunk under consideration. For example, step 407 may include using a trained machine learning algorithm (e.g., from Figure 3A ) to analyze one patient vascular trunk at a time and use features of the vascular tree from a machine learning algorithm to infer the healthy lumen diameter at the trunk under consideration.

[0097] In one embodiment, step 409 may include inferring a healthy lumen diameter for the vessel segment. For example, an exemplary embodiment may include extracting patient-specific indicators (including features of step 405), omitting one site at a time, and mapping a database of these indicators to a stored healthy lumen diameter (e.g., from a training set such as Figure 3A More specifically, the machine learning algorithm may include random forest regression to predict the healthy lumen diameter of the patient. As previously described, the random forest may use an ensemble of decision trees, where each decision tree may be composed of a random subset of features and training data. Each decision tree may map an input feature vector to a continuous variable, and the values ​​from the decision trees may be pooled together and averaged to calculate the healthy lumen diameter (d p )’s final predicted value.

[0098] In one embodiment, step 411 may include determining the LNS of the vessel segment. For example, the local lumen diameter (d l ) and the predicted healthy lumen diameter. The LNS or stenosis percentage is estimated as α = (1-d p / d l ) x 100%.

[0099] In a study (Sankaran S., Schaap M., Hunley SC, Min JK, Taylor CA, Grady L. (2016) HALE:Healthy Area of ​​Lumen Estimation for Vessel Stenosis Quantification .In: Ourselin S., Joskowicz L., Sabuncu M., Unal G., Wells W. (eds.) Medical Image Computing and Computer-Assisted Intervention--MICCAI2016. MICCAI 2016. Lecture Notes in Computer Science, vol 9902. Springer, Cham), method 400 achieves a correlation coefficient of 0.947 with a mean absolute error of 0.15 mm for predicting lumen diameter in healthy parts. In addition, the method can have an operating point sensitivity / specificity of 90% / 85% for detecting stenosis. Compared with QCA data, the mean absolute error of stenosis percentage on a group of diseased patients can be reduced from 31% in anisotropic kernel regression to 14% in the present method 300.

[0100] A general approach using patient-specific geometric features (including vessel volume, lumen area, vessel length, and combinations of the resulting features) can be used when estimating healthy lumen diameter using a random decision forest regressor. This regressor can be used for vascular systems with different types of diseases such as acute, diffuse, open, and bifurcated. The reference kernel regression-based approach described earlier in this disclosure can be based on local patient-specific data. Such a regression approach can capture areas of focal stenosis. Optional or additional regression approaches can interpret population data and capture other disease morphologies. Yet another approach, such as a method for detecting diffuse lesions, can include a population-based machine learning approach, in which an output indicator (LNS) can provide an indication of the presence or absence of diffuse lesions.

[0101] The current system and method can be used with any lumen segmentation algorithm. Depending on the application, the current system and method can be used, for example, with an automated lumen segmentation algorithm for on-site estimation of percent stenosis or with a semi-automated approach offline or in a core laboratory setting. The current system and method can provide an accurate QCT assessment tool, which can involve coupling an accurate lumen segmentation algorithm with an accurate algorithm for estimating percent stenosis. Such an assessment tool can work well with QCA and better than anisotropic kernel regression for the same lumen segmentation.

[0102] An additional embodiment may include prohibiting the calculation of features on sites that may be diseased. For example, such an embodiment may include an iterative algorithm where once a site is identified as diseased when training a machine learning algorithm, it may not be used in the estimation of features at other sites. Embodiments may also include higher order metrics, such as area gradients.

[0103] Other embodiments of the invention will be apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. It is intended that the specification and examples be considered exemplary only, with the true scope and spirit of the invention being indicated by the following claims.

Claims

1. A computer-implemented method for identifying a lumen diameter of a patient's vasculature, the method comprising: include: receiving a dataset comprising one or more lumen segments of known healthy blood vessel segments of a plurality of individuals; For each known healthy blood vessel segment, extract lumen features of a portion of the vascular system upstream of the healthy blood vessel segment; For each known healthy blood vessel segment, extract lumen features of a portion of the vascular system downstream of the healthy blood vessel segment; and calculating a population-based healthy lumen diameter based on the extracted lumen features of the portion of the vasculature upstream of the healthy vessel segment and based on the extracted lumen features of the portion of the vasculature downstream of the healthy vessel segment; The patient-specific healthy lumen diameter is calculated based on the population-based healthy lumen diameter by: receiving a lumen segment of the patient's vasculature; determining a location of the patient's vasculature; as well as The determined population-based healthy lumen diameter is used to determine a patient-specific healthy lumen diameter for the region of the patient's vasculature.

2. The method according to claim 1, further comprising: include: A patient-specific luminal stenosis score is calculated based on the patient-specific healthy luminal diameter.

3. The method according to claim 1, further comprising: include: Each of the lumen segments of the known healthy vessel segment is divided into sub-units, wherein one of the sub-units corresponds to a determined location of the patient's vasculature.

4. The method according to claim 1, further comprising: include: An estimate of the fractional flow reserve is generated, an estimate or sensitivity of the fractional flow reserve estimate is generated, or a model is generated based on the patient-specific luminal stenosis score.

5. The method according to claim 1, further comprising: include: An estimate of the fractional flow reserve is generated, an estimate or sensitivity of the fractional flow reserve estimate is generated, or a model is generated based on a healthy lumen diameter specific to the patient.

6. The method of claim 1, wherein the one or more lumen characteristics include mean, maximum and minimum lumen area, volume and length.

7. A system for identifying the lumen diameter of a patient's vascular system, the system include: A processor configured to perform a method comprising: receiving a dataset comprising one or more lumen segments of known healthy blood vessel segments of a plurality of individuals; For each known healthy blood vessel segment, extract lumen features of a portion of the vascular system upstream of the healthy blood vessel segment; For each known healthy blood vessel segment, extract lumen features of a portion of the vascular system downstream of the healthy blood vessel segment; and calculating a population-based healthy lumen diameter based on the extracted lumen features of the portion of the vasculature upstream of the healthy vessel segment and based on the extracted lumen features of the portion of the vasculature downstream of the healthy vessel segment; The patient-specific healthy lumen diameter is calculated based on the population-based healthy lumen diameter by: receiving a lumen segment of the patient's vasculature; determining a location of the patient's vasculature; as well as The determined population-based healthy lumen diameter is used to determine a patient-specific healthy lumen diameter for the region of the patient's vasculature.

8. The system of claim 7, wherein the system is further configured to: A patient-specific luminal stenosis score is calculated based on the patient-specific healthy luminal diameter.

9. The system of claim 7, wherein the system is further configured to: Each of the lumen segments of the known healthy vessel segment is divided into sub-units, wherein one of the sub-units corresponds to a determined location of the patient's vasculature.

10. The system of claim 7, wherein the system is further configured to: An estimate of the fractional flow reserve is generated, an estimate or sensitivity of the fractional flow reserve estimate is generated, or a model is generated based on the patient-specific luminal stenosis score.

11. The system of claim 7, wherein the system is further configured to: An estimate of the fractional flow reserve is generated, an estimate or sensitivity of the fractional flow reserve estimate is generated, or a model is generated based on a healthy lumen diameter specific to the patient.

12. The system of claim 7, wherein the one or more lumen characteristics include mean, maximum and minimum lumen area, volume and length.

13. A non-transitory computer readable medium for use on a computer system, the medium containing computer executable program instructions for performing a method of identifying a lumen diameter of a patient's vasculature, the method include: receiving a dataset comprising one or more lumen segments of known healthy blood vessel segments of a plurality of individuals; For each known healthy blood vessel segment, extract lumen features of a portion of the vascular system upstream of the healthy blood vessel segment; For each known healthy blood vessel segment, extract lumen features of a portion of the vascular system downstream of the healthy blood vessel segment; and calculating a population-based healthy lumen diameter based on the extracted lumen features of the portion of the vasculature upstream of the healthy vessel segment and based on the extracted lumen features of the portion of the vasculature downstream of the healthy vessel segment; The patient-specific healthy lumen diameter is calculated based on the population-based healthy lumen diameter by: receiving a lumen segment of the patient's vasculature; determining a location of the patient's vasculature; as well as The determined population-based healthy lumen diameter is used to determine a patient-specific healthy lumen diameter for the region of the patient's vasculature.

14. The non-transitory computer readable medium of claim 13, wherein the method further comprises: include: A patient-specific luminal stenosis score is calculated based on the patient-specific healthy luminal diameter.