U.S. patent application number 17/238735 was filed with the patent office on 2021-08-19 for machine learning system for assessing heart valves and surrounding cardiovascular tracts.
This patent application is currently assigned to Stenomics, Inc.. The applicant listed for this patent is Stenomics, Inc.. Invention is credited to Michael A. SINGER.
Application Number | 20210257097 17/238735 |
Document ID | / |
Family ID | 1000005564726 |
Filed Date | 2021-08-19 |
United States Patent
Application |
20210257097 |
Kind Code |
A1 |
SINGER; Michael A. |
August 19, 2021 |
MACHINE LEARNING SYSTEM FOR ASSESSING HEART VALVES AND SURROUNDING
CARDIOVASCULAR TRACTS
Abstract
A machine learning system for evaluating at least one
characteristic of a heart valve, an inflow tract, an outflow tract
or a combination thereof may include a training mode and a
production mode. The training mode may be configured to train a
computer and construct a transformation function to predict an
unknown anatomical characteristic and/or an unknown physiological
characteristic of a heart valve, inflow tract and/or outflow tract,
using a known anatomical characteristic and/or a known
physiological characteristic the heart valve, inflow tract and/or
outflow tract. The production mode may be configured to use the
transformation function to predict the unknown anatomical
characteristic and/or the unknown physiological characteristic of
the heart valve, inflow tract and/or outflow tract, based on the
known anatomical characteristic and/or the known physiological
characteristic of the heart valve, inflow tract and/or outflow
tract.
Inventors: |
SINGER; Michael A.;
(Belmont, CA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Stenomics, Inc. |
Belmont |
CA |
US |
|
|
Assignee: |
Stenomics, Inc.
Belmont
CA
|
Family ID: |
1000005564726 |
Appl. No.: |
17/238735 |
Filed: |
April 23, 2021 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
16050613 |
Jul 31, 2018 |
11024426 |
|
|
17238735 |
|
|
|
|
15923032 |
Mar 16, 2018 |
10943698 |
|
|
16050613 |
|
|
|
|
15138922 |
Apr 26, 2016 |
9953272 |
|
|
15923032 |
|
|
|
|
14680892 |
Apr 7, 2015 |
9424531 |
|
|
15138922 |
|
|
|
|
14511018 |
Oct 9, 2014 |
9092743 |
|
|
14680892 |
|
|
|
|
61894814 |
Oct 23, 2013 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
A61B 5/7253 20130101;
A61B 5/7267 20130101; A61B 34/10 20160201; G16H 50/50 20180101;
G06N 20/00 20190101; G06N 5/04 20130101; A61B 5/7275 20130101; G16H
50/20 20180101; G06F 17/11 20130101 |
International
Class: |
G16H 50/20 20060101
G16H050/20; A61B 5/00 20060101 A61B005/00; G06N 5/04 20060101
G06N005/04; G06N 20/00 20060101 G06N020/00; A61B 34/10 20060101
A61B034/10; G16H 50/50 20060101 G16H050/50; G06F 17/11 20060101
G06F017/11 |
Claims
1. A computer-implemented method for evaluating at least one
characteristic of a heart valve, an inflow tract, an outflow tract
or a combination thereof the method comprising: (a) predicting,
with a transformation function on a computer, an unknown anatomical
characteristic of at least one of a training heart valve, a
training inflow tract or a training outflow tract, using at least
one of a training known anatomical characteristic or a training
known physiological characteristic of the at least one training
heart valve, training inflow tract or training outflow tract,
wherein said at least one of said training heart valve, said
training inflow tract, and said training outflow tract includes at
least one of said training heart valve, a training coronary vessel,
a training Valsalva sinus, a training sinotubular junction, and a
training ascending aorta; (b) using a production mode of a system
on the computer to direct the transformation function and one or
more feature vectors, to predict unknown anatomical characteristic
of at least one production heart valve, production inflow tract or
production outflow tract, based on at least one of patient specific
production known anatomical characteristic or production known
physiological characteristic of at least one said production heart
valve, said production inflow tract or said production outflow
tract, wherein said at least one of said production heart valve,
said production inflow tract, said production outflow tract
includes at least one of said production heart valve, a production
Valsalva sinus, a production sinotubular junction, and a production
ascending aorta, to generate at least one or more quantities of
interest.
2. A method as in claim 1 further maintaining, in said at least one
or more feature vectors on the computer, the at least one patient
known anatomical characteristic or patient known physiological
characteristic of the at least one production heart valve,
production inflow tract or production outflow tract.
3. A method as in claim 2, further comprising using the computer to
calculate an approximate blood flow through the at least one
production heart valve, production inflow tract or production
outflow tract.
4. A method as in claim 2, further comprising using the computer to
store in at least one of said one or more feature vectors
quantities associated with an approximate blood flow through the at
least one production heart valve, production inflow tract or
production outflow tract.
5. A method as in claim 2, further comprising using the computer to
perturb the at least one patient known anatomical characteristic or
patient known physiological characteristic of the at least one
production heart valve, production inflow tract or production
outflow tract stored in the at least one of said one or more
feature vectors.
6. A method as in claim 5, further comprising using the computer to
calculate a new approximate blood flow through the at least one
production heart valve, production inflow tract or production
outflow tract with the perturbed at least one patient known
anatomical characteristic or patient known physiological
characteristic.
7. A method as in claim 5, further comprising using the computer to
store quantities associated with a new approximate blood flow
through the perturbed at least one production heart valve,
production inflow tract or production outflow tract in said one or
more feature vectors.
8. A method as in claim 7, further comprising using the computer to
repeat said perturbing and said storing to create at least one of
(a) one or more feature vectors and (b) one or more quantity
vectors.
9. A method as in claim 1, further comprising the production mode
using the computer to apply the transformation function to the one
or more feature vectors.
10. A method as in claim 9, further comprising the production mode
using the computer to generate the one or more quantities of
interest.
11. A method as in claim 10, further comprising the production mode
using the computer to store the one or more quantities of
interest.
12. A method as in claim 11, further comprising the production mode
using the computer to process the quantities of interest to provide
data for use in at least one of evaluation, diagnosis, prognosis,
risk, treatment and treatment planning related to at least one of
the production heart valve, production inflow tract, and production
outflow tract.
13. A method as in claim 12 further comprising using said data to
at least one of (1) guide clinical decision-making, (2) provide
predictive information about disease progression, (3) provide
information for risk stratification, (4) patient monitoring, (5)
conducting sensitivity analyses, (6) evaluating an anatomic
scenario, (7) evaluating a physiologic scenario, (8) evaluating a
hemodynamic scenario, (9) estimating response to therapy, and (10)
developing understanding of cardiac health.
14. A method as in claim 1 wherein said computer-implemented method
includes a computed tomography device.
15. A method as in claim 1 wherein said computer-implemented method
includes a magnetic resonance imaging device.
16. A method as in claim 1 wherein said computer-implemented method
includes an ultrasound imaging device.
17. A method as in claim 1 wherein said system comprises a Doppler
device.
18. A method as in claim 1 wherein said system comprises an
electrophysiologic device.
19. A method as in claim 1 wherein said system comprises clinical
instruments.
20. A method as in claim 1, further comprising the production mode
using the computer to process the quantities of interest to provide
data for use in at least one of evaluation, diagnosis, prognosis,
risk, treatment and treatment planning related to at least one of
the production heart valve, production inflow tract, and production
outflow tract.
21. A method as in claim 1, further comprising the production mode
using the computer to provide data for use in at least one of the
construction and execution of a computer-based model of at least
one of cardiac anatomy and physiology.
22. A method as in claim 1 further comprising: (a) using a training
mode of a system on the computer to train said computer and
construct said transformation function based upon a plurality of
images to predict said unknown anatomical characteristic of at
least one of said training heart valve, said training inflow tract
or said training outflow tract, using at least one of said training
known anatomical characteristic or said training known
physiological characteristic of the at least one training heart
valve, training inflow tract or training outflow tract; (b) wherein
said training known anatomical characteristic or said training
known physiological characteristic of the at least one training
heart valve, training inflow tract or training outflow tract
includes at least one of size, shape, and flow characteristics.
23. A method as in claim 22 wherein said transformation function is
based upon at least one morphological simplification that exploits
underlying geometric features.
24. A method as in claim 22 wherein said training known anatomical
characteristic or said training known physiological characteristic
of the at least one training heart valve, training inflow tract or
training outflow tract characterizes calcification.
25. A method as in claim 24 wherein said characterizes
calcification includes location of calcification.
26. A method as in claim 24 wherein said characterizes
calcification includes extent of calcification.
27. A method as in claim 24 wherein said characterizes
calcification includes size of calcification.
28. A method as in claim 24 wherein said characterizes
calcification includes degree of calcification.
29. A method as in claim 21, further comprising the training mode
using the computer to store in said at least one of said one or
more feature vectors which are representative of the at least one
training known anatomical characteristic or training known
physiological characteristic of the at least one training heart
valve, training inflow tract or training outflow tract.
30. A method as in claim 29, further comprising the training mode
using the computer to calculate an approximate blood flow through
the at least one training heart valve, training inflow tract or
training outflow tract.
31. A method as in claim 29, further comprising the training mode
using the computer to store in at least one or more feature vectors
quantities associated with an approximate blood flow through the at
least one training heart valve, training inflow tract or training
outflow tract.
32. A method as in claim 29, further comprising the training mode
using the computer to perturb the at least one training known
anatomical characteristic or training known physiological
characteristic of the at least one training heart valve, training
inflow tract or training outflow tract stored in at least one of
said one or more feature vectors.
33. A method as in claim 32, further comprising the training mode
using the computer to calculate a new approximate blood flow
through the at least one training heart valve, training inflow
tract or training outflow tract with the perturbed at least one
training known anatomical characteristic or training known
physiological characteristic.
34. A method as in claim 32, further comprising the training mode
using the computer to store quantities associated with a new
approximate blood flow through the perturbed at least one training
heart valve, training inflow tract or training outflow tract.
35. A method as in claim 34, further comprising the training mode
using the computer to repeat the perturbing and storing to create a
set of feature vectors and quantity vectors.
36. A method as in claim 1, further comprising the production mode
using the computer to generate the one or more quantities of
interest that includes hemodynamic data.
37. A method as in claim 1, further comprising the production mode
using the computer to generate the one or more quantities of
interest that includes patient data.
38. A method as in claim 1, further comprising the production mode
using the computer to generate the one or more quantities of
interest that includes the location of the at least one production
heart valve, production inflow tract or production outflow
tract.
39. A method as in claim 1, further comprising the production mode
using the computer to generate the one or more quantities of
interest that includes qualitative information of the at least one
production heart valve, production inflow tract or production
outflow tract.
40. A method as in claim 1, further comprising the production mode
using the computer to generate the one or more quantities of
interest that includes quantitative information of the at least one
production heart valve, production inflow tract or production
outflow tract for use in evaluating valvular anatomy.
41. A method as in claim 1, further comprising the production mode
using the computer to generate the one or more quantities of
interest that includes quantitative information of the at least one
production heart valve, production inflow tract or production
outflow tract for use in evaluating physiology.
42. A method as in claim 1 wherein said transformation function
characterizes discretized surface elements.
43. A method as in claim 1 wherein said transformation function
characterizes discretized volume elements.
44. A method as in claim 1 wherein any of said unknown anatomic
characteristic is an unknown geometrical anatomic
characteristic.
45. A method as in claim 1 wherein said transformation is a
multi-dimensional transformation function.
46. A method as in claim 1 wherein said transformation is a
one-dimensional transformation function.
47. A method as in claim 1, further comprising using the computer
to perform the following steps: (a) receiving patient-specific data
selected from the group consisting of anatomic data, physiologic
data, and hemodynamic data; (b) generating a digital model of the
at least one production heart valve, production inflow tract or
production outflow tract, based on the received data; (c)
discretizing the digital model; (d) applying boundary conditions to
at least one inflow portion and at least one outflow portion of the
digital model; and (e) initializing and solving mathematical
equations of blood flow through the digital model.
48. A method as in claim 47, further comprising the computer
storing quantities and parameters that characterize at least one of
an anatomic state or a physiologic state of the digital model and
the blood flow.
49. A method as in claim 47, further comprising the computer
perturbing at least one of an anatomic parameter or a physiologic
parameter that characterizes the digital model.
50. A method as in claim 49, further comprising the computer at
least one of re-discretizing or re-solving the mathematical
equations with the at least one anatomic parameter or physiologic
parameter.
51. A method as in claim 50, further comprising the computer
storing quantities and parameters that characterize at least one of
the anatomic state or the physiologic state of the perturbed model
and blood flow.
52. A method as in claim 1, further comprising the production mode
using the computer to receive one or more feature vectors.
53. A method as in claim 52, further comprising the production mode
using the computer to apply the transformation function to the
feature vectors.
54. A method as in claim 53, further comprising the production mode
using the computer to generate one or more quantities of
interest.
55. A method as in claim 54, further comprising the production mode
using the computer to process the quantities of interest to provide
data for use in at least one of evaluation, diagnosis, prognosis,
treatment or treatment planning related to a heart in which the at
least one production heart valve, production inflow tract or
production outflow tract resides.
Description
CROSS-REFERENCE TO RELATED APPLICATIONS
[0001] This application is a continuation of U.S. patent
application Ser. No. 16/050,613, filed Jul. 31, 2018, which is a
continuation of U.S. patent application Ser. No. 15/923,032, filed
Mar. 16, 2018, now U.S. Pat. No. 10,943,698, issued Mar. 9, 2021,
which is a continuation of U.S. patent application Ser. No.
15/138,922, filed Apr. 26, 2016, now U.S. Pat. No. 9,953,272,
issued Apr. 24, 2018, which is a continuation of U.S. patent
application Ser. No. 14/680,892, filed Apr. 7, 2015, now U.S. Pat.
No. 9,424,531, issued Aug. 23, 2016, which is a continuation of
U.S. patent application Ser. No. 14/511,018, filed Oct. 9, 2014,
now U.S. Pat. No. 9,092,743, issued Jul. 28, 2015, which claims
priority to U.S. Provisional Patent Application No. 61/894,814,
filed on Oct. 23, 2013. The full disclosures of the above-listed
patent applications are hereby incorporated by reference
herein.
TECHNICAL FIELD
[0002] The present disclosure relates generally to the fields of
machine learning, computer modeling and simulation, and computer
aided design. More specifically, the disclosure relates to
computer-based machine learning systems and methods for
constructing and executing models of cardiac anatomy and
physiology. These models may be used for therapeutic, treatment,
and/or diagnostic purposes.
BACKGROUND OF THE INVENTION
[0003] Cardiovascular disease is the leading cause of death in the
United States and claims the lives of more than 600,000 Americans
each year. According to the American Heart Association (AHA), more
than five million Americans are diagnosed with heart valve disease
each year, and diseases of the aortic and mitral valves are the
most prevalent. Combined, aortic and mitral valve diseases affect
more than five percent of the U.S. population.
[0004] The proper assessment and diagnosis of heart valve operation
and the condition of surrounding cardiovascular tracts are
essential for ensuring high quality patient care. To this end,
several imaging modalities may be used to inspect the condition and
function of heart valves and the surrounding vasculature.
Transthoracic and transesogophogeal echocardiography, for example,
use ultrasound technology to create two- and/or three-dimensional
images of heart valves and the surrounding inflow/outflow tracts
(e.g., left ventricular outflow tract, ascending aorta). Further,
computed tomography (CT) and magnetic resonance imaging (MRI) may
also be used.
[0005] All imaging modalities have strengths and weaknesses that
may limit their ability to provide a complete and comprehensive
assessment of anatomic and/or physiologic condition. The spatial
resolution of echocardiographic images, for example, may inhibit a
detailed analysis of functional operation, especially for highly
calcified heart valves. Computed tomography may provide higher
resolution images than echocardiography, but CT imaging studies are
more costly and expose patients to radiation that is potentially
harmful. In addition, contrast agents, which may be highly
nephrotoxic and may be associated with alterations in renal
function, are often used during CT examinations. Hence, new and
novel methods that enable an accurate anatomic and physiological
assessment of heart valves and the surrounding vasculature, while
not exposing patients to excessive risks or prohibitive costs, are
desirable.
[0006] Patients diagnosed with symptomatic and clinically
significant heart valve abnormalities may be candidates for
valvular repair or replacement. When repair or replacement is
indicated, an accurate and complete understanding of valvular
anatomy is essential to ensure a favorable outcome. In addition,
the anatomic and physiologic characteristics of the inflow and
outflow tracts that surround the heart valve(s) must also be
understood.
[0007] New methods for assessing the anatomic and/or physiologic
condition of native and prosthetic heart valves and the surrounding
inflow/outflow tracts should enable more accurate and precise
treatment planning. These new methods may complement and/or work in
conjunction with existing methods, or they may stand alone.
Regardless, such technologies must provide clear and demonstrable
benefits to the physician(s) who treat patients with heart valve
disease and/or diseases of the surrounding cardiac tracts. Further,
new technologies must not expose patients to excessive medical
risks and should be cost effective.
[0008] Therefore, to improve diagnostic and treatment capabilities,
it is desirable to have a system for quickly and accurately
assessing the physiological function, condition, and morphology of
heart valves and the surrounding inflow/outflow tracts, which
thereby enables the proper diagnosis of heart valve disease and, if
warranted, facilitates treatment planning.
DESCRIPTION OF RELATED ART
[0009] There are many academic and industrial research groups that
use computer modeling and simulation to analyze flow through heart
valves. Historically, valvular hemodynamic analyses have focused on
the aortic heart valve and have employed methods of computational
fluid dynamics (CFD) to provide detailed insight into the blood
flow surrounding the aortic valve. These insights have then been
used to facilitate the design and construction of heart valves with
desirable hemodynamic properties that maximize functionality and
durability while minimizing the potentially fatal risks of valvular
malfunction and adverse physiological response
[0010] In recent years, hemodynamic modeling of heart valves has
included both surgically implanted and transcatheter prostheses,
but the focus of most studies remains the aortic valve. With the
rapidly expanding clinical deployment of transcatheter aortic heart
valves, modeling and simulation results have helped understand and
characterize the unique hemodynamic challenges of transcatheter
designs compared to traditional surgical implantation of aortic
valves. In particular, computer modeling may be used to quantify
downstream flow effects in the aortic arch and leaflet stresses,
which impact device efficacy, robustness, durability, and
longevity.
[0011] To date, all computer modeling and simulation studies of
heart valves have been focused on evaluating and improving
prosthetic valve design and function.
BRIEF SUMMARY OF THE PRESENT INVENTION
[0012] The machine learning system and method described in this
disclosure facilitates the diagnosis and treatment of heart valve
disease and diseases of the surrounding inflow/outflow tracts.
Further, the system and method facilitate the evaluation and
assessment of valvular repair and/or prosthetic performance in
patients who have undergone heart valve treatment. In addition to
using routine physiological and geometric data gathered through
two- and/or three-dimensional imaging studies, the machine learning
system may also incorporate hemodynamic data into the construction
and utilization of an accurate geometric and functional
understanding from which to assess valvular condition and
function.
[0013] In one aspect, a machine learning system for evaluating at
least one characteristic of a heart valve, an inflow tract and/or
an outflow tract may include a training mode and a production mode.
The training mode may be configured to train a computer and
construct a transformation function to predict an unknown
anatomical characteristic and/or an unknown physiological
characteristic of a heart valve, an inflow tract and/or an outflow
tract, using a known anatomical characteristic and/or a known
physiological characteristic of the heart valve, inflow tract
and/or outflow tract. The production mode may be configured to use
the transformation function to predict the unknown anatomical
characteristic and/or the unknown physiological characteristic of
the heart valve, inflow tract and/or outflow tract, based the known
anatomical characteristic and/or the known physiological
characteristic of the heart valve, inflow tract and/or outflow
tract.
[0014] In some embodiments, the training mode is configured to
compute and store in a feature vector the known anatomical
characteristic and/or known physiological characteristic of the
heart valve, inflow tract and/or outflow tract. In some
embodiments, the training mode is configured to calculate an
approximate blood flow through the heart valve, inflow tract and/or
outflow tract. In some embodiments, the training mode is further
configured to store quantities associated with the approximate
blood flow through the heart valve, inflow tract and/or outflow
tract. Optionally, the training mode may be further configured to
perturb the at least one known anatomical characteristic or known
physiological characteristic of the heart valve, inflow tract
and/or outflow tract stored in the feature vector. In some
embodiments, the training mode may be further configured to
calculate a new approximate blood flow through the heart valve,
inflow tract and/or outflow tract with the perturbed known
anatomical characteristic and/or known physiological
characteristic. In some embodiments, the training mode may be
further configured to store quantities associated with the new
approximate blood flow through the perturbed heart valve, inflow
tract and/or outflow tract. In some embodiments, the training mode
may be further configured to repeat the perturbing, calculating and
storing steps to create a set of feature vectors and quantity
vectors and to generate the transformation function.
[0015] In one embodiment, the training mode may be further
configured to perform a method, involving: receiving
patient-specific data including anatomic data, physiologic data
and/or hemodynamic data; generating a digital model of the at least
one heart valve, inflow tract or outflow tract, based on the
received data; discretizing the digital model; applying boundary
conditions to at least one inflow portion and at least one outflow
portion of the digital model; and initializing and solving
mathematical equations of blood flow through the digital model. In
some embodiments, the method may further involve storing quantities
and parameters that characterize an anatomic state and/or a
physiologic state of the digital model and the blood flow. In some
embodiments, the method may further involve perturbing an anatomic
parameter and/or a physiologic parameter that characterizes the
digital model. In another embodiment, the method may further
involve re-discretizing and/or re-solving the mathematical
equations with the anatomic parameter and/or physiologic parameter.
In another embodiment, the method may further involve storing
quantities and parameters that characterize the anatomic state
and/or the physiologic state of the perturbed model and blood
flow.
[0016] In some embodiments, the production mode may be configured
to receive one or more feature vectors. In some embodiments, the
production mode may be configured to apply the transformation
function to the feature vectors. In some embodiments, the
production mode may be configured to generate one or more
quantities of interest. In some embodiment, the production mode may
be configured to store the quantities of interest. In some
embodiments, the production mode may be configured to process the
quantities of interest to provide data for use in at least one of
evaluation, diagnosis, prognosis, treatment or treatment planning
related to a heart in which the heart valve resides.
[0017] In another aspect, a computer-implemented machine learning
method for evaluating at least one characteristic of a heart valve,
an inflow tract, and/or an outflow tract may involve training a
computer by using a training mode of a machine learning system to
construct a transformation function to predict an unknown
anatomical characteristic and/or an unknown physiological
characteristic a heart valve, an inflow tract and/or an outflow
tract, using a known anatomical characteristic and/or a known
physiological characteristic of the heart valve, inflow tract
and/or outflow tract. The method may also involve using a
production mode of the machine learning system to direct the
transformation function to predict the unknown anatomical
characteristic and/or the unknown physiological characteristic of
the heart valve, inflow tract and/or outflow tract, based on the
known anatomical characteristic and/or the known physiological
characteristic of the heart valve, inflow tract and/or outflow
tract.
[0018] In some embodiments, the method may further involve using
the training mode to compute and store in a feature vector the
known anatomical characteristic and/or known physiological
characteristic of the heart valve, inflow tract and/or outflow
tract. In some embodiments, the method may further involve using
the training mode to calculate an approximate blood flow through
the heart valve, inflow tract and/or outflow tract. In some
embodiments, the method may further involve using the training mode
to store quantities associated with the approximate blood flow
through the heart valve, inflow tract and/or outflow tract. In some
embodiments, the method may further involve using the training mode
to perturb the known anatomical characteristic and/or known
physiological characteristic of the heart valve, inflow tract
and/or outflow tract stored in the feature vector. In some
embodiments, the method may further involve using the training mode
to calculate a new approximate blood flow through the heart valve,
inflow tract and/or outflow tract with the perturbed known
anatomical characteristic and/or known physiological
characteristic. In some embodiments, the method may further involve
using the training mode to store quantities associated with the new
approximate blood flow through the perturbed heart valve, inflow
tract and/or outflow tract. In some embodiments, the method may
further involve using the training mode to repeat the perturbing,
calculating and storing steps to create a set of feature vectors
and quantity vectors and to generate the transformation
function.
[0019] In some embodiments, the method may further involve using
the training mode to perform the following steps: receiving
patient-specific data selected from the group consisting of
anatomic data, physiologic data, and hemodynamic data; generating a
digital model of the at least one heart valve, inflow tract or
outflow tract, based on the received data; discretizing the digital
model; applying boundary conditions to at least one inflow portion
and at least one outflow portion of the digital model; and
initializing and solving mathematical equations of blood flow
through the digital model. In some embodiments, the method may
further involve storing quantities and parameters that characterize
an anatomic state and/or a physiologic state of the digital model
and the blood flow. In some embodiments, the method may further
involve perturbing an anatomic parameter and/or a physiologic
parameter that characterizes the digital model. In some
embodiments, the method may further involve re-discretizing or
re-solving the mathematical equations with the at least one
anatomic parameter or physiologic parameter. In some embodiments,
the method may further involve storing quantities and parameters
that characterize the anatomic state and/or the physiologic state
of the perturbed model and blood flow.
[0020] In some embodiments, the method may further involve
receiving one or more feature vectors with the production mode. In
some embodiments, the method may further involve using the
production mode to apply the transformation function to the feature
vectors. In some embodiments, the method may further involve using
the production mode to generate one or more quantities of interest.
In some embodiments, the method may further involve using the
production mode to store the quantities of interest. In some
embodiments, the method may further involve using the production
mode to process the quantities of interest to provide data for use
in evaluation, diagnosis, prognosis, treatment and/or treatment
planning related to a heart in which the heart valve, inflow tract
and/or outflow tract resides.
[0021] In another aspect, a non-transitory computer readable medium
for use on a computer system may contain computer-executable
programming instructions for performing a method for evaluating at
least one characteristic of a heart valve, an inflow tract, an
outflow tract or a combination thereof. The method may include any
of the features and/or aspects described above.
[0022] In various other aspects, this disclosure describes various
method embodiments. Examples of such method embodiments include: A
method of using data analysis and/or machine learning to construct
a transformation function to compute the anatomic and/or
physiologic state of at least one heart valve and/or the
corresponding inflow/outflow tracts; A method of using computer
modeling and simulation and/or clinical data to generate a set of
feature vectors that are used as input into a machine learning
algorithm; A method of using machine learning to assess anatomy
and/or physiology of at least one heart valve and/or the
corresponding inflow/outflow tracts, comprising using
patient-specific data derived from one or more interventional or
non-interventional methods and/or results generated by computer
modeling and simulation; A method of using machine learning to
assess the anatomy and/or physiology of at least one heart valve
and/or the corresponding inflow/outflow tracts, comprising using
patient-specific data derived from one or more interventional or
non-interventional methods to perform sensitivity and uncertainly
analyses; A method of using machine learning to assess the anatomy
and/or physiology of at least one heart valve and/or the
corresponding inflow/outflow tracts, comprising using
patient-specific data derived from one or more interventional or
non-interventional methods to aid in the diagnosis, assessment
and/or prognosis of a diseased state; and A method of using machine
learning to assess the anatomy and/or physiology of at least one
heart valve and/or the corresponding inflow/outflow tracts,
comprising using patient-specific data derived from one or more
interventional or non-interventional methods to aid in the planning
of prosthetic heart valve implantation.
[0023] These and other aspects and embodiments will be described in
further detail below, in reference to the attached drawing
figures.
BRIEF DESCRIPTION OF THE DRAWINGS
[0024] FIG. 1 is a block diagram of a machine learning system,
according to one embodiment;
[0025] FIG. 2 is a flow diagram outlining a modeling and simulation
method for a training portion of a machine learning system,
according to one embodiment;
[0026] FIG. 3 is a flow diagram outlining execution of a training
portion of a machine learning system, according to one
embodiment;
[0027] FIG. 4 is a flow diagram outlining execution of a production
portion of a machine learning system, according to one
embodiment;
[0028] FIG. 5 is a perspective view of a simplified geometric
model, based on patient-specific anatomic parameters, of an aortic
valve and surrounding cardiac inflow and outflow vessels, according
to one embodiment;
[0029] FIG. 6 is a perspective view of a simplified geometric model
with the computational surface mesh, based on patient-specific
anatomic parameters, of the aortic valve and the surrounding
cardiac inflow and outflow vessels, according to one embodiment;
and
[0030] FIGS. 7A-7D are perspective views of various representative
polyhedra used to discretize the interior volume of the geometric
model, according to various embodiments.
DETAILED DESCRIPTION OF PREFERRED EMBODIMENT
[0031] This disclosure describes machine learning systems and
methods that qualitatively and quantitatively characterize anatomic
geometry and/or physiology of a heart valve, one or more inflow
tracts of a heart valve, and/or one or more outflow tracts of a
heart valve. Throughout this disclosure, reference may be made to
characterizing or evaluating a heart valve. In all embodiments,
such characterization, evaluation, etc. may be performed on a heart
valve, one or more inflow tracts of a heart valve, and/or one or
more outflow tracts of a heart valve. For enhanced readability of
the description, however, the phrase "heart valve" may simply be
used, rather than repeating "a heart valve, one or more inflow
tracts of a heart valve, and/or one or more outflow tracts of a
heart valve" in each instance. Any embodiment described for use in
evaluating a heart valve may additionally or alternatively be used
to evaluate one or more inflow tracts of a heart valve and/or one
or more outflow tracts of a heart valve. The various embodiments
described herein may be applied to any single heart valve, a
combination of multiple heart valves, and/or combinations of one or
more heart valves and one or more coronary blood vessels. Although
occasional references may be made to one specific heart valve,
inflow tract, or outflow tract, these specific references should
not be interpreted as limiting the scope of this disclosure. For
example, the aortic heart valve is used throughout this disclosure
as a specific example of a prototypical heart valve. Illustration
of the systems and methods via the example of the aortic heart
valve, however, is not intended to limit the scope of the computer
modeling and simulation systems and methods disclosed herein.
[0032] Referring to FIG. 1 and according to one embodiment, a
machine learning system 30 may include two modes: a training mode
32 and a production mode 34. The two modes 32, 34 may be embodied
in a computer system and/or a computer readable medium. The system
30 may execute the two modes in series, where the training mode 32
is executed first, and the production mode 34 is executed second.
The training mode 32 may be configured to develop analytical
capabilities in a computer system that enable the computer system
to predict unknown anatomic and/or physiologic characteristics of
one or more heart valves and/or the surrounding inflow/outflow
tracts. These predictive capabilities may be developed by the
analysis and/or evaluation of known anatomic and/or physiologic
characteristics of one or more heart valves and/or the surrounding
inflow/outflow tracts. Using a collection of known anatomic and/or
physiologic characteristics, a computer may be "trained" to predict
various unknown anatomic and/or physiologic characteristics. The
abstract mapping that transforms a set of known characteristics
into one or more predictions of unknown characteristics may be
referred to as the "transformation function." In some embodiments,
the training mode 32 may be configured to construct the
transformation function.
[0033] The production mode 34 of the machine learning system 30 may
use the transformation function to predict anatomic and/or
physiologic characteristics that are unknown from a collection of
anatomic and/or physiologic characteristics that are known. Hence,
during execution of the production mode 34, input into the
transformation function may be a set of known anatomic and/or
physiologic characteristics (e.g., the same anatomic and/or
physiologic characteristics used during the training mode 32). The
output of the transformation function may be one or more anatomic
and/or physiologic characteristics that were previously
unknown.
[0034] The training mode 32 and production mode 34 may be
implemented in a number of different ways in various alternative
embodiments. One embodiment of a method for implementing the
training mode 32 and production mode 34 of a machine learning
system is described in more detail immediately below. This is only
one exemplary embodiment, however, and should not be interpreted as
limiting the scope of the machine learning system 30 as described
above.
Training Mode:
[0035] During the training mode 32 of the machine learning system
30, anatomic and/or physiologic data may be acquired that
characterize the state and operation of a heart valve and its
corresponding inflow/outflow tracts. These data may be collected
through one or more acquisition methods, including, for example,
analysis of radiological images, analysis of echocardiographic
images, Doppler and/or electrophysiologic signals, clinical
instruments (e.g., blood pressure gauge, stethoscope), and computer
modeling/simulation. Referring to the aortic valve as an example,
anatomic and/or physiologic characterization parameters may
include, for example:
[0036] flow characteristics (e.g., velocities, velocity gradients,
pressures, pressure gradients, turbulence intensity, shear stress)
at single or multiple location(s) within the left ventricular
outflow tract (LVOT), valsalva sinuses (VS), sinotubular junction
(SJ), ascending aorta (AA) or vasculature surrounding one or more
heart valve(s);
[0037] approximations to flow, flow properties or flow
characteristics via simplified and/or analytical models (e.g., pipe
flow, orifice flow);
[0038] size and/or shape characteristics at single or multiple
location(s) within the LVOT, VS, SJ, AA, or surrounding
vasculature, e.g., diameter, eccentricity, cross-sectional area,
axial length, length of major axis, length of minor axis, geometric
gradient(s);
[0039] height, shape, lateral profile, thickness, degree of
calcification, location of calcification, angular size, angular
separation, radial length, tip sharpness, rigidity, flexibility,
movement, tissue properties, overlap, and/or attachment angle(s) of
one or more valve leaflets;
[0040] location, attachment angles, and/or sizes of one or more
coronary arteries;
[0041] geometric orifice area and/or estimated orifice area of the
valve;
[0042] size, shape, location, density, composition, and/or extent
of vascular calcification;
[0043] stroke volume and/or cardiac output;
[0044] blood pressure, heart rate, and/or hematocrit of the
patient; and
[0045] age, height, weight, body mass index, race, and/or gender of
the patient.
[0046] Referring to FIG. 2, one embodiment of a method for
implementing the training mode 32 of the machine learning system 30
is illustrated. In this embodiment, the training mode 32 of the
machine learning system 30 is coupled with a modeling and
simulation system (not shown), which may provide input data for the
machine learning system 30. Hence, the modeling and simulation
system may operate in conjunction with the machine learning system
30, in that it may provide anatomic and/or physiologic data to the
machine learning system 30. These data may serve as the foundation
from which the machine learning system 30 learns to perform the
desired task(s).
[0047] A first step of the embodiment described in FIG. 2 may
involve importing patient-specific geometric, anatomic,
physiologic, and/or hemodynamic data into the computer system 100.
A second step may involve constructing a (possibly parameterized)
geometric model using the imported data 200. One embodiment of a
geometric model 10 is illustrated in FIG. 5.
[0048] As illustrated in FIG. 5, in one embodiment, the geometric
model 10 may be a multi-dimensional digital representation of the
relevant patient anatomy, which may include at least one heart
valve 12 (the aortic valve in one embodiment), at least a portion
of an inflow vessel 14 (or "inflow tract"), and at least a portion
of an outflow vessel 16 (or "outflow tract") of the valve 12. The
model may also include one or more ventricles and/or atria of the
heart or a portion thereof and/or one or more coronary vessels or a
portion thereof. The geometric model is created from
patient-specific anatomical, geometric, physiologic, and/or
hemodynamic data. In some embodiments, the model may be created
using exclusively imaging data. Alternatively, the model may be
created using imaging data and at least one clinically measured
flow parameter. Imaging data may be obtained from any suitable
diagnostic imaging exam(s), such as those listed above. Clinically
measured flow parameters may be obtained from any suitable test(s),
such as those listed above.
[0049] The model 10 may also contain at least one inflow boundary
and at least one outflow boundary, through which blood flows in and
out of the multi-dimensional model 10, respectively. These inflow
and outflow boundaries denote finite truncations of the digital
model 10 and are not physically present in a patient. The digital
geometric model 10 may be created using methods of applied
mathematics and image analysis, such as but not limited to image
segmentation, machine learning, computer aided design, parametric
curve fitting, and polynomial approximation. In some embodiments, a
hybrid approach, which combines a collection of geometric modeling
techniques, may also be utilized. The final, multi-dimensional
model 10 provides a digital surrogate that captures the relevant
physical features of the anatomic topology under consideration and
may contain one or more morphological simplifications (e.g.,
symmetry, smoothing) that exploit the underlying geometric features
of the patient-specific valvular and vascular system being
considered.
[0050] Referring again to FIG. 1, following the construction of the
digital model 200, the modeling and simulation portion of the
machine learning system may discretize the surface and volume of
the model into a finite number of partitions 300. These individual
and non-overlapping partitions, called "elements," may facilitate
the application and solution of the physical laws of motion that
govern blood flow through the geometric model. The set of surface
and volume elements used to discretize the model, collectively
referred to as the "mesh," transform the continuous geometric model
into a set of mesh points and edges, where each element point in
the mesh has discrete x, y, and z spatial coordinates, and each
element edge is bounded by two mesh points and has a finite
length.
[0051] An illustration of a representative mesh 21 that discretizes
the surface of a geometric model 20 is shown in FIG. 6. The
geometric model 20, in this embodiment, includes an aortic valve
22, inflow tract 24 and outflow tract 26. This illustration of the
model 20 is used to show the mesh 21 and is intended for exemplary
purposes only.
[0052] The shape of the surface elements created by the modeling
and simulation portion of the machine learning system may take the
form of any closed polygon, but the surface mesh typically contains
a collection of triangles, convex quadrilaterals or a combination
thereof. Referring to FIGS. 7A-7D, volume elements may be created
by the modeling and simulation system and are used to fill the
interior of the model completely. Each volume element may take the
form of any closed polyhedron, but the volume mesh (i.e., the set
of volume elements) typically contains a collection of tetrahedra
(FIG. 7A), hexahedra (FIG. 7B), pyramids (FIG. 7C), wedges (FIG.
7D), or a combination thereof. The surface and volume mesh
densities, which determine the spatial resolution of the discrete
model, may vary in space and time. The local densities of the
surface and volume meshes may depend on the complexity of the local
topology of the underlying geometric model: more complex local
topology may require higher spatial resolution, and therefore a
higher mesh density, to resolve than local regions of less complex
topology.
[0053] The modeling and simulation portion of the machine learning
method may use CFD to simulate blood flow through the discretized
geometric model. Blood may be represented as a Newtonian or
non-Newtonian fluid, and blood flow may be represented physically
by the conservation of mass, momentum, and energy (or a combination
thereof) and mathematically by the fluid flow equations (e.g.,
continuity, Navier-Stokes equations) with appropriate initial and
boundary conditions. The boundary conditions may be a function of
time and/or space. Initial and boundary conditions may be
determined from empirical or heuristic relationships, clinical
data, mathematical formulas or a combination thereof, and the model
boundaries may be rigid or compliant or a combination thereof. The
mathematical equations and corresponding initial and boundary
conditions may be solved using conventional mathematical
techniques, which include analytical or special functions,
numerical methods (e.g., finite differences, finite volumes, finite
elements, spectral methods), methods of machine learning or a
hybrid approach that combines various aspects of the methods
listed.
[0054] As a next step in the modeling and simulation portion of the
machine learning method, and referring again to FIG. 2, boundary
conditions may be applied to a discrete patient model 400. The
boundary flow conditions may be obtained from patient-specific
clinical measurements (e.g., pulse wave Doppler echocardiography,
continuous wave Doppler echocardiography, MRI), in which case they
may be prescribed to the model in a manner that is consistent with
clinical observations and measurements. In addition, inflow and
outflow boundary conditions may be prescribed to compensate for
underlying psychological or medical conditions such as pain,
anxiety, fear, anemia, hyperthyroidism, left ventricular systolic
dysfunction, left ventricular hypertrophy, hypertension or
arterial-venous fistula, which may produce clinically misleading
results upon which medical evaluations, diagnostics, treatment
planning or treatment(s) may be based.
[0055] With continued reference to FIG. 2, following the
initialization of the blood flow equations, the equations are
solved, and hemodynamic quantities of interest are computed 500 by
the modeling and simulation system, which may be a component of the
training mode 32 of the machine learning system 30. The hemodynamic
quantities of interest computed by the modeling and simulation
system may include, for example, the flow velocity at one or more
points in the computational domain, velocity gradients, pressure,
pressure gradients, shear stress, the wall shear stress at
location(s) on the heart valve, etc.
[0056] Following the solution of the mathematical equations and
computation of the quantities of interest, the anatomic and
physiologic parameters that are inputs into the modeling and
simulation system, collectively referred to as "features," may be
assembled into a vector 600. This vector of anatomic and
physiologic features is referred to as a "feature vector." As an
illustrative example, numerical quantities contained in a feature
vector may include some or all of the parameters (or features)
outlined above, e.g., LVOT diameter, LVOT velocity, LVOT cross
sectional area, height of each valvular leaflet, thickness of each
valvular leaflet, diameter of the ascending aorta, etc. The
corresponding hemodynamic quantities of interest, which may be
computed from the CFD simulation from an anatomic model that may be
characterized by features in the feature vector, may also assembled
into a vector, which may be referred to as the "quantity of
interest vector." The quantity of interest vector may include, for
example, wall shear stress, pressure, pressure gradients, velocity,
velocity gradients, and/or shear at various locations throughout
the model, etc. Both the feature and quantity of interest vectors
may then be saved for use during other steps of the machine
learning process. Note that a feature vector and the corresponding
quantity of interest vector may have different lengths. In
addition, entries within the feature and quantity of interest
vector may be obtained from different mechanisms (e.g., clinical
data, numerical simulations, estimated approximation). Nonetheless,
each feature vector is associated with a quantity of interest
vector and vice versa.
[0057] Referring to FIG. 2, a next step in the method may involve
modifying (or "perturbing") the digital model and/or flow condition
to represent perturbed anatomic and/or physiologic conditions 700.
As an example of an anatomic perturbation, one valve leaflet may be
retracted to increase the geometric orifice area of the valve. As
an example of a physiologic perturbation, the inflow velocity
through the LVOT may be increased or decreased.
[0058] As illustrated in FIG. 2, following modification(s) to the
anatomic and/or physiologic conditions 700, steps 300-700 of the
modeling and simulation portion of the machine learning system may
be repeated 800, until a desired number of feature vectors and the
corresponding quantities of interest vectors are obtained. Note
that each iteration of steps 300-700 produces a new feature vector
and a new quantity of interest vector. Though one or more entries
within the feature and/or quantity of interest vector may change
with each iteration of steps 300-700, the representation and length
of each vector remains the same. That is, each digital model is
represented by the same characteristics and the same number of
characteristics, and this collection of characteristics is
contained within the feature vector. Further, the corresponding
quantities of interest for each digital model are the same. The
sets of feature and quantity of interest vectors may then be stored
on digital media.
[0059] In some embodiments, and referring now to FIG. 3, a machine
learning method may involve applying machine learning algorithms to
a collection of feature and quantity of interest vectors from the
method described above and illustrated in FIG. 2. The collection of
feature and quantity of interest vectors may first be imported into
machine learning software 900. The machine learning software may
then apply one or more analysis or machine learning algorithms
(e.g., decision trees, support vector machines, regression,
Bayesian networks, random forests) to the set of feature and
quantity of interest vectors 1000. Following the application of
machine learning algorithm(s), a transformation function is
constructed 1100. This transformation function may serve as a
mapping between the one or more features contained within a feature
vector and the one or more quantities of interest computed from the
modeling and simulation portion of the machine learning system.
Hence, the input into the transformation function is a feature
vector, and the output of the transformation function is a quantity
of interest vector. To test the accuracy of the transformation
function created by the machine learning algorithm, for example,
one of the feature vectors used to create the transformation
function may be used as input into the transformation function. The
expected output from the transformation function is the
corresponding quantity of interest vector, though the quantity of
interest output vector may not be reproduced exactly by the
transformation function. The transformation function may be stored
on digital media for use, for example, during the production mode
of the machine learning system 1200.
[0060] Following construction of the transformation function by the
analysis and machine learning algorithm(s), functioning of the
training mode 32 of the machine learning system 30, as described in
the present embodiment, may be complete. Subsequently, the
transformation function may be used in the production mode 34 of
the machine learning system 30.
Production Mode:
[0061] The production mode 34 of the machine learning system 30 may
be used after the training mode 32. The production mode 34 may be
configured to compute quantity of interest vectors rapidly and
accurately by applying the transformation function to a variety of
feature vectors. In some but not all cases, these feature vectors
might have been used to construct the transformation function.
[0062] Referring now to FIG. 4, in one embodiment, the production
mode 34 of the machine learning system 30 may first be used to
import the transformation function and one or more feature vectors
1300, which contain the same set of features used during the
training mode 32. The feature vectors used during the production
mode 34 may or may not have been used during the training mode to
construct the transformation function, and therefore the
transformation function may not have been constructed with data
contained within these feature vectors. The number of features
within each feature vector and the quantities represented by each
feature within each feature vector, however, are the same as those
used to construct the transformation function.
[0063] The transformation function may then be applied to the one
or more feature vectors 1400. Hence, the inputs to the
transformation function during the production mode 34 of the
machine learning system 30 may be one or more feature vectors, and
the output from the transformation function may be a vector that
contains the quantities of interest. The quantity of interest
vector outputted from the transformation function may then be
stored 1500, e.g., on digital media.
[0064] The quantities of interest contained within the quantity of
interest vector may include qualitative and/or quantitative
geometric and hemodynamic information. These data may be further
analyzed and assessed through various mechanisms of post-processing
to reveal patient-specific anatomic and/or physiologic and/or
hemodynamic information that may aid in the diagnosis, treatment,
and/or treatment planning of a patient. These qualitative and
quantitative data may also be used to guide clinical
decision-making and/or provide predictive information about disease
progression or risk stratification.
[0065] Quantities of interest and/or data derived from the machine
learning system 30 may be delivered to physicians, who may use
these data for clinical decision-making. Delivery of
patient-specific information to physicians may occur via integrated
or stand-alone software systems, numerical data, graphs, charts,
plots, verbal discussions, written correspondence, electronic
media, etc. or a combination thereof. These data may then be used
by an individual physician or by a team of physicians to develop a
complete, comprehensive, and accurate understanding of patient
cardiac health and to determine whether or not medical treatment is
warranted. If medical treatment is warranted, results from the
machine learning system 30 may be used to guide clinical
decision-making. By way of example, specific ways in which output
from the machine learning system 30 may be incorporated into the
clinical management of cardiac patients include:
[0066] analysis of heart valve operation, including diagnosing the
severity, functional significance, and clinical response to
abnormal heart valve operation;
[0067] patient-specific selection, sizing, and positioning of
prosthetic heart valves, including surgical, transcatheter, and
valve-in-valve treatments; and
[0068] patient monitoring and/or follow-up.
[0069] The list of applications outlined above is for example
purposes only, and the list is not intended to be exhaustive.
[0070] The machine learning system 30 may provide a fast and
accurate virtual framework for conducting patient-specific
sensitivity analyses. Such analyses may assess the relative impacts
of geometric and/or hemodynamic changes to the anatomic,
physiologic, and/or hemodynamic state of a patient; these state
changes may then be assessed for functional and clinical
significance thereby estimating patient response to therapy (or
lack thereof), disease progression, and/or patient-specific risk
stratification. Sensitivity analyses may be performed, for example,
by applying the transformation function, which is computed during
the training mode 32 of the machine learning system 30, to multiple
feature vectors that describe variations of specific anatomic
and/or physiologic features of the patient. Although construction
of the transformation function during the training mode 32 is
likely best to include feature vectors that are similar to those
used during a sensitivity analysis, it is important to note that
the transformation function may not require re-computation during a
sensitivity analysis study. Hence, the machine learning system 30
may enable a rapid evaluation of numerous anatomic, physiologic,
and/or hemodynamic scenarios that run in a virtual environment
without exposing patients to any medical risks. Results from the
plethora of transformation function evaluations conducted during a
sensitivity analysis may be aggregated and presented to physicians
for clinical decision-making. Further, results from sensitivity
analyses may also be used in conjunction with uncertainty analyses
to, for example, assess global and/or local uncertainties of
anatomic, physiologic, and/or hemodynamic results produced by the
machine learning system 30.
[0071] The machine learning system 30 enables planning of heart
valve replacement therapy and the selection of optimal valve
deployment. For example, executing the machine learning system 30
described herein provides an accurate assessment of anatomic,
physiologic, and/or hemodynamic consideration for valvular
deployment and function, e.g., size, deployment mechanism,
deployment angle. Hence, the machine learning system 30 and methods
for using it provide a complete framework that enables the accurate
assessment of anatomic structure in relation to native and
prosthetic heart valves and their corresponding inflow/outflow
tracts. This information may be used by physicians to make clinical
decisions regarding patient treatment of heart valve disease as to
maximize the benefits to each patient.
[0072] Although the above description highlights a number of
embodiments and examples, the present invention extends beyond the
specifically disclosed embodiments to other alternative embodiments
and/or uses of the invention and modifications and equivalents
thereof. Thus, the scope of the present invention should not be
limited by the particular disclosed embodiments described above,
but should be determined only by a fair reading of the claims that
follow.
* * * * *