Patent Application
20240120095
The inventive subject matter is generally directed towards systems and methods for facilitating risk assessment of adverse health outcomes. In particular, embodiments of the invention relate to AI-based systems and methods for determining risks of adverse cardiovascular outcomes based, at least in part, on the use of a chest computerized tomography (CT) scan and AI-enabled analysis of the signals detected in the CT scan based on coronary artery calcifications (CAC) and cardiac chambers volumetry. Additional features include one or more of cardiac ejection fraction, pericardial fat, thoracic aortic calcification, aortic valve calcification, mitral valve calcification, pulmonary artery hypertension, emphysema, chronic obstructive pulmonary disease (COPD), fatty liver disease and bone mineral density.
It is estimated that over 15 million individuals die every year from cardiovascular disease (CVD), with the majority being unaware of their risk until they experience a symptomatic life threatening condition which is often too late to reverse. Coronary Artery Disease (CAD) specifically and atherosclerotic cardiovascular disease (ASCVD) generally are the largest part of CVD death worldwide. Early detection of high risk ASCVD is crucial, particularly for asymptomatic individuals, as common symptoms such as chest pain and shortness of breath often do not manifest until a few minutes before a sudden cardiac death.
CAC scoring has been used to prove the presence of CAD and to facilitate predicting the risk of serious coronary events. A CAC score facilitates risk prediction and is more predictive than any other single biomarker. Adding a CAC score to traditional risk factors improves the Area Under the Curve (AUC) for ASCVD risk prediction.
The Agatston method quantifies the amount of calcium in coronary arteries. The Agatston score is a measure of coronary artery calcification obtained from a computed tomography (CT) scan of the heart. It is calculated by adding the scores of individual calcified plaques in the coronary arteries, considering the peak intensity of a calcified signal and area of each calcified plaque as defined by HU over 130. The resulting score represents the overall burden of calcium in the coronary arteries and serves as an indicator of an individual's CAD risk. A higher Agatston score indicates a greater amount of calcium buildup in the coronary arteries, correlating with a higher risk of ASCVD. The score can range from 0 (no detectable calcium) to several thousand, with scores above 400 indicating a high risk of coronary artery disease. The Agatston score is one of several methods used to assess a person's risk of heart disease, and is often combined with other factors, such as age, gender, and family history, to determine overall risk.
There is an unmet need for early detection of pre-symptomatic CVD, and for affordable, reliable, and scalable technology. While CAD can be assessed by a coronary artery calcium (CAC) scan which is a far more powerful predictor of ASCVD death than blood pressure and lipid levels, it is far from perfect. For example, in the NIH sponsored multi-ethnic study of atherosclerosis (MESA) over 15 years follow up of 6814 people, 99 people out of those with Agatston CAC=0 had an adverse coronary event.
Short-Term versus Long-Term Risk Prediction. CVD has been the primary cause of death and healthcare costs in the US for decades. Every year over 600,000 first-time heart attacks unexpectedly hit asymptomatic Americans. Currently less than 3% of US adults aged 20-79 years have an optimal cardiovascular risk factors profile defined as: total cholesterol <200 mg/dL (5.17 mmol/L), blood pressure <120/<80 mm Hg, non-smoker, body mass index (BMI)<25 kg/m2, fasting plasma glucose <100 mg/dL (5.56 mmol/L). Nonetheless, the awareness on CVD risk factors is above 95% meaning that almost all US adults are aware of the risk associated with these risk factors. Clearly, new strategies are needed.
Since the pioneering Framingham Heart Study in 1960s introduced CVD risk factors, preventive cardiology has focused on long-term CVD risk prediction. Currently, physicians tell their patients that based on their risk factors (age, gender, blood pressure, cholesterol, diabetes, smoking, etc.) their risk of developing CVD in the next 10 years is X. The median and mean for X are 2.7% and 5.2% respectively. Although such a long-term risk assessment is necessary, it is not enough. It does not trigger immediate preemptive actions and cannot detect asymptomatic patients who are vulnerable to a near-term CVD event. A layman's analogy to this scenario would be a TV weather broadcaster announcing that, over the next 10 years, a catastrophic hurricane will hit your area. Such an announcement would hardly change behaviors. However, when the weatherman displays a hurricane eye coming your way in the near future, it can cause immediate preemptive actions. A medical analogy would be finding a tumor in a cancer patient that gets serious attention and triggers immediate interventions to improve outcomes. Having a near-term predictive tool in cardiology might cause a paradigm shift resulting in developing new treatments. Because of the multi-factorial nature of CVD, the solution is preferably a multi-factorial prediction model based on several variables.
For CAD, screening and diagnostic tools like CAC score and coronary CT angiography are known. To predict atrial fibrillation (AF) and heart failure (HF), CHARGE-AF and brain natriuretic peptide (BNP) are used. CHARGE AF is an epidemiological risk calculator, not suitable for an individualized risk assessment and monitoring, while BNP is more precise but not specific to, for example, left atrial and ventricular function.
AF is the most common sustained arrhythmia and carries an increased risk of stroke and cardiovascular mortality. In the United States, at least 3 to 6 million people experience AF. The number of AF patients is expected to rise to over 12 million cases by 2030, imposing a significant economic burden of $260 million in healthcare expenses. In Europe, AF among those older than 55 years is projected to reach 14 million by 2040. It is estimated that by 2050, at least 72 million people in Asia will be diagnosed with AF, and about 3 million with AF-related strokes. This presents a public health crisis especially for the growing elderly population in the coming decades.
Medicare services costs are significantly higher among AF patients than non-AF patients; therefore, early treatment is critical to limit the disease burden imposed by AF. The adverse social and public health effects of HF are even worse. It is estimated that by 2030, more than 8 million people in the United States will have HF. And the total direct medical costs of HF are expected to rise from $21 billion to $53 billion.
The 5-year survival rates of AF are of concern. Without proper treatment, 51% of AF patients will die within five years. Although the economic burden posed by AF and HF are critical considering the increasing healthcare costs, early detection tools and preventive interventions for pre-AF and pre-HF patients are currently unavailable. One report shows that 96,860 strokes occurred within 1 year among patients with AF, with an associated total direct lifetime cost of nearly $8 billion. Of these costs, $2.6 billion in direct costs are incurred during the first year after the stroke.
HF poses an even greater threat to the American healthcare system. Given the rising rates of hospitalization and re-hospitalization, HF is associated with a significant cost burden. Approximately 1% to 2% of the total US health care budget is spent on HF, and half of that is attributable to late diagnosis leading to inpatient admissions for HF. This challenge presents a great opportunity to make an impact on the healthcare system by early detection and interventions of subclinical HF and AF. Currently, BNP and CHARGE-AF are the only available tools for early detection of high-risk patients for AF and HF. A combination of high-risk CHARGE-AF and a 7-day ECG patch has been reported. CHARGE-AF is an epidemiological risk calculator that can be useful as a population-based measure, but it is not preferred as applied to individual patients as needed in a physician's office. A more direct assessment is preferred such as by imaging the cardiac chambers where AF happens.
It is known to use manual measurements of left ventricle chamber in a single slice to predict HF. However, rapid and accurate acquisition of whole heart volume parameters is challenging. Even though semi-automated delineation and quantification of cardiovascular structures can be useful in CT images, currently known methods still require a significant degree of manual modifications, which is time-consuming and may increase inter-/intra-observer variability.
Approximately half of all adults aged 50 and older are at risk of bone fractures due to osteoporosis or osteopenia. From 1990 to 2019, global deaths and disability-adjusted life-years (DALYs) attributable to low bone mineral density (BMD) increased by 112% and 94%, respectively. Proper treatment can prevent about half of the repeat fractures related to osteoporosis. Likewise, most osteopenia cases can be prevented and reversed with appropriate healthcare. However, a significant number of individuals with osteopenia and osteoporosis remain unaware of their bone loss. A BMD test serves as the sole method for early determination of a suitable treatment plan to prevent further bone loss and future fractures.
Dual-energy x-ray absorptiometry (DEXA) can be used for assessing BMD, but less than 20% of the population who should get BMD test undergoes one. DEXA is limited by its 2D planar technique making it unable to distinguish between cortical and trabecular bone, leading to underestimated bone loss, especially in overweight individuals. Further, osteoporosis is usually asymptomatic prior to the fracture event resulting in fractures becoming the dominant clinical manifestation. Consequently, osteoporosis remains an underdiagnosed and undertreated condition associated with high morbidity and mortality.
Lung cancer is also a major disease affecting humans. It is the second most common diagnosed cancer in the United States, and it accounts for the greatest number of cancer deaths in both men and women worldwide. Early detection of lung cancer has been challenging and it was not until 2011 with the release of data from the National Lung Cancer Screening Trial (NLST) that a screening test for lung cancer was demonstrated to reduce lung-cancer specific mortality. More specifically, this trial demonstrated that use of low dose computed tomography (LDCT) for lung cancer screening resulted in a significant reduction rate in lung cancer mortality, so that LDCT is now the standard of care for lung cancer screening.
Currently, no screening tool is available for detecting individuals at high risk of AF and/or HF. Despite the critical need no biomarker is currently available to identify individuals at high risk for AF or HF and their complications. An ideal biomarker would help to detect individuals at high risk for both AF and HF. In view of the above, there is a long-felt need in the healthcare industry to improve predictive risk assessment of adverse health outcomes in health care settings. Embodiments of the systems and methods disclosed here address these and other needs in the relevant art.
Data integration projects face challenges bringing together the right expertise. There are pipeline behaviors that can be common across multiple pipelines. However, under known methods, typically pipeline behaviors are not developed once and made reusable by different pipelines. Some embodiments disclosed herein address the need for scale where dozens of projects may be concurrently active.
In one aspect, the invention is directed to an AI-based method of assessing the risk of adverse health outcomes. In one embodiment, the method involves providing an artificial intelligence system (AIS) configured to analyze a set of computed tomography (CT) scan images; inputting a set of CT scan images into the AIS for analysis; receiving an AI analysis from the AIS, the AI analysis including: (a) a coronary artery calcium (CAC) score based, at least in part, on a modified Agatston score; (b) a measurement of mean Hounsfield units (HU) per plaque adjusted by plaque volume; (c) a measurement of plaque quantity per coronary artery and per patient; (d) a measurement of the number of coronary arteries with one or more plaque; (e) a measurement of location of plaques from proximal, to aortic root, to distal coronary artery; (f) a volume of at least one cardiac chamber; wherein the AI analysis is based at least in part on the set of CT scan images; and determining, based at least in part on a computerized risk calculator, the AI analysis, and known risk factors, a patient's risk for an adverse health outcome.
In some embodiments, the known risk factors include one or more of age, gender, ethnicity, smoking, blood pressure, blood lipids, blood glucose, hemoglobin A1C, brain natriuretic peptide, presence of diabetes, endothelial dysfunction, and cardiometabolic syndromes. In certain embodiments, the AI analysis further includes a measurement of at least one of: plaque shape, distance from other plaques, distribution of plaques mean HU per coronary artery, aortic valve calcification, aortic wall calcification, aortic diameter, pulmonary arteries diameter, left atrium, left ventricle, right atrium, right ventricle, thoracic bone mineral density, pericardial fat, and liver fat, intra thoracic fat, emphysema, lung nodules, subcutaneous fat, and thoracic muscle mass.
In certain embodiments, determining a patient's risk for an adverse health outcome further comprises configuring the risk calculator for specialized assessment of different adverse outcomes selected from the group including coronary heart disease, congestive heart failure, atrial fibrillation, left ventricular hypertrophy, stroke, chronic obstructive pulmonary disease, cardiovascular death, and all-cause mortality. In one embodiment, determining a patient's risk for an adverse health outcome further comprises configuring the risk calculator to facilitate specialized short-term risk assessment of adverse health outcomes to provide an alert of an imminent risk, wherein the short-term can be days, weeks, and/or up to 12 months.
In some embodiments, the AI analysis further includes a measurement of plaque calcifications defined by HU≥100. In certain embodiments, the set of CT scan images includes images obtained from contrast-enhanced cardiac CT scans. In one embodiment, the method can further include monitoring the amount of calcification in the coronary arteries, cardiac and aortic valve areas overtime to track progression or regression of aortic valve calcification and gauging response to treatments. In some embodiments, the AI analysis can further include a display of coronary arteries vasa vasorum density around coronary arteries by mapping the Hounsfield units where the areas corresponding to highest vasa vasorum density have the highest Hounsfield units.
In another aspect the invention concerns an AI-based system for facilitating risk assessment of adverse health outcomes. In one embodiment, the system can include a device configured to facilitate acquiring and storing a set of CT scan images; and an AI-enabled analyzer configured to generate an analysis based, at least in part, on said set of scan images. The analysis preferably includes a coronary artery calcium (CAC) score based, at least in part, on an Agatston score; a measurement of plaque density; a measurement of plaque quantity; one or more locations of plaque; and a volume of at least one cardiac structure. The system can further include a computer-enabled risk calculator configured to determine, based at least in part on the analysis, a particular patient's risk for an adverse health outcome.
In some embodiments, the analysis can further include a measurement of thoracic aortic calcification and a measurement of bone mineral density. In certain embodiments, the computer-enabled risk calculator is further configured to determine the patient's risk for an adverse health outcome based, at least in part, on at least one known risk factor selected from the group: age, gender, ethnicity, smoking, blood pressure, blood lipids, blood glucose, hemoglobin A1C, brain natriuretic peptide, presence of diabetes, endothelial dysfunction, and cardiometabolic syndromes.
In certain embodiments, the computer-enabled risk calculator is configured to determine a particular patient's risk for an adverse health outcome selected from: coronary heart disease, congestive heart failure, atrial fibrillation, stroke, chronic obstructive pulmonary disease, cardiovascular death, and all-cause mortality. In yet other embodiments, the computer-enabled risk calculator is further configured to facilitate specialized short-term risk assessment of adverse health outcomes to provide an alert of an imminent risk; the short-term can be days, weeks, and/or up to 12 months. In certain embodiments, the AI analysis further includes a measurement of plaque calcifications defined by HU≥100.
Yet another aspect of the invention relates to an AI-enabled method of measuring cardiac ejection fraction (EF). In one embodiment, the method involves providing an artificial intelligence system (AIS) configured to estimate cardiac chambers volumes based on non-contrast ECG-gated CT scan images of the heart; providing a first image comprising a non-contrast, ECG-gated CT scan image of the heart acquired during end-diastolic period; providing a second image comprising a non-contrast, ECG-gated CT scan image of the heart acquired during end-systolic period; wherein the first and second measurements are based, at least in part, on the first and second images; and calculating the difference in left ventricular volume between the first and second images; and determining an EF measurement based, at least in part, on the first and second measurements.
In some embodiments, the first image is acquired during isovolumetric contraction, and wherein the second image is acquired during isovolumetric relaxation. In certain embodiments, the method can further include receiving from the AIS a third measurement comprising a measurement of LV wall volume and total heart volume to calculate total heart volume changes between end-diastolic period and end-systolic period.
Another aspect of the invention is directed to an AI-enabled system (AIS) configured to facilitate detecting individuals at high risk of future adverse health outcomes. In one embodiment, the system includes an AI-based module configured to receive and analyze 2-D chest X-ray images; and a computerized calculator configured to detect individuals at high risk of one or more of: atrial fibrillation, heart failure, and stroke; wherein the AI-based module is configured to perform an analysis based, at least in part, on detecting in the X-ray images characteristics of an enlarged left atrium, enlarged left atrial appendage, enlarged right atrium, dilated pulmonary arteries, dilated pulmonary veins, and/or enlarged right and left ventricles.
Yet another aspect of the invention concerns an AI-enabled system (AIS) configured to improve the clinical utility of coronary artery calcium (CAC) scans. In one embodiment, the system includes an AI-based module configured to extract from the CAC scans cardiac chambers volumetry data and an Agatston CAC score; and a computerized CVD risk calculator configured to provide a minimum Net Reclassification Index (NRI) of over 0.1 for risk assessment of individuals at risk of future heart failure, atrial fibrillation, stroke, CVD related death, and all-cause mortality; and wherein the computerized CVD risk calculator is configured to provide the NRI based, at least in part, on the cardiac chambers volumetry data and the Agatston CAC score.
Another aspect of the invention is directed to an AI-based method of assessing the risk of adverse health outcomes. In one embodiment, the method includes providing a first artificial intelligence system (AIS1) configured to analyze a set of computed tomography (CT) scan images; inputting a set of CT scan images into AIS1 for analysis; receiving an AI analysis from AIS1, the AI analysis including: a coronary artery calcium (CAC) score based, at least in part, on an Agatston score; a measurement of mean, median, standard deviation, and range of Hounsfield units (HU) per plaque adjusted by plaque volume; a measurement of plaque quantity per coronary artery and per patient; a measurement of the number of coronary arteries with one or more plaque; a measurement of location of plaques from proximal to the aortic root, to distal parts of a coronary artery; a volume of at least one cardiac chamber; wherein the AI analysis is based at least in part on the set of CT scan images; providing a second AIS (AIS2) configured to classify data associated with coronary plaques and/or cardiac chambers based on the output of AIS1; and determining, based at least in part on a computerized risk calculator, the AIS1 analysis, the classification performed by AIS2, and known risk factors, a patient's risk for an adverse health outcome.
One more aspect of the invention concerns an AI-based method of assessing and monitoring changes in the risk for adverse health outcomes. In one embodiment, the method includes providing an artificial intelligence system (AIS) configured to analyze a set of computed tomography (CT) scan images and to output an analysis comprising a cardiovascular risk score based on five or more measurement from the group comprising: a measurement of the Agatston calcium score; a measurement of mean, median, standard deviation, and range of Hounsfield units (HU) per plaque adjusted by plaque volume; a measurement of plaque quantity per coronary artery and per patient; a measurement of the number of coronary arteries with one or more plaque; a measurement of location of plaques from proximal to the aortic root, to distal parts of a coronary artery; a measurement of the length of the longest axis of a plaque; a measurement of proximity of plaques to each other within a coronary artery; a volume of one or more cardiac chambers comprising left atrium, left ventricle, right atrium, right ventricle, and left ventricular wall; a measurement of bone mineral density using average HU in trabecular bones; a measurement of lung nodules; and a measurement of lung emphysema score. In some embodiments, the AI analysis output can be inputted into a computerized risk calculator configured to determine, based on AIS analysis output and one or more of known risk factors, a patient's risk for an adverse health outcome.
Additional features and advantages of the embodiments disclosed herein will be set forth in the detailed description that follows, and in part will be clear to those skilled in the art from that description or recognized by practicing the embodiments described herein, including the detailed description which follows, the claims, as well as the appended drawings.
Both the foregoing general description and the following detailed description present embodiments intended to provide an overview or framework for understanding the nature and character of the embodiments disclosed herein. The accompanying drawings are included to provide further understanding and are incorporated into and constitute a part of this specification. The drawings illustrate various embodiments of the disclosure, and together with the description explain the principles and operations thereof.
A more complete understanding of the embodiments, and the attendant advantages and features thereof, will be more readily understood by references to the following detailed description when considered in conjunction with the accompanying drawings wherein:
The specific details of the single embodiment or variety of embodiments described herein are set forth in this application. Any specific details of the embodiments are used for demonstration purposes only, and no unnecessary limitation or inferences are to be understood therefrom.
Before describing in detail exemplary embodiments, it is noted that the embodiments reside primarily in combinations of components related to the system. Accordingly, the device components have been represented where appropriate by conventional symbols in the drawings, showing only those specific details that are pertinent to understanding the embodiments of the present disclosure so as not to obscure the disclosure with details that will be readily apparent to those of ordinary skill in the art having the benefit of the description herein.
HeartLung Corporation makes available several AI-enabled tools: coronary artery calcium score (AutoCAC) using non-contrast chest CT scans; bone mineral density (AutoBMD) measurements; and cardiac structure volumetry (AutoChamber). AutoCAC can also provide density, quantity, and location of plaques. AutoChamber can also provide information associated with pericardial fat, thoracic aorta calcium, aorta, and pulmonary aorta sizing. HeartLung's technology can work on new scans and existing scans.
There are imaging features of coronary calcification that can enable improving the risk stratification particularly in patients with calcium score of 1 to 100. These features include calcium density (scattered versus dense), shape of calcification (sharp versus blunt edges), heterogeneity (focal or segmental around the calcified spot) and distribution (distal and proximal location). Calcium density is not properly used in Agatston scoring. The Agatston score is weighted based on the maximum intensity Hounsfield unit (HU) of pixels corresponding to a calcified plaque in the coronary arteries, but it does not take into account the average intensity (HU) of the calcified plaque neither it considers the distribution and standard deviation of HU. For a given plaque area, the Agatston score is increased 2, 3, or 4-fold based on reaching higher calcium density HU cut points which elevates the calcium score. However, densely calcified coronary plaques may pose a lower CVD risk than less densely calcified plaques because they are presumably more stable and have less lipid core. The importance of differences in the degree of atherosclerotic plaque calcification is studied in randomized trials with CAC and statin therapy. Older and more established plaques are more likely to calcify. Conversely, newer plaques have less calcium, greater lipid content and are more likely to be vulnerable. Deep learning techniques can facilitate identification of specific imaging features that can differentiate individuals and reclassify those at risk of having cardiovascular events with low calcium score (1-100) or zero calcium score.
One of the embodiments of this inventive matter teaches a significant improvement over Agatston Score and is here referred to as the modified Agatston method. Unlike the Agatston score, in one embodiment, the modified Agatston method is weighted by the number of plaques, mean HU intensity (not maximum HU intensity that Agatston uses) adjusted by number of pixels, location of plaque, and/or vessels. An Agatston score 1-100 is a weakness of coronary artery calcium scoring using the prior art Agatston method. This is similar to the weakness of the Intermediate Risk category of existing ASCVD risk factors-based scoring method (ASCVD score), which is widely known to be inaccurate in the Intermediate Risk category where most heart attacks happen instead of in the High Risk category. In the case of the existing Agatston Score, about 25% of adverse events happen in the 1-100 Agatston score category. The modified Agatston facilitates identifying individuals who are actually High Risk in this CAC 1-100 population and should not take a CAC 1-100 lightly. This is a major improvement over the existing Agatston CAC score.
In addition to enhanced CAC features, certain embodiments of the invention disclosed here measure quantitative radiomic features of other components in a chest CT scan including cardiac chambers volume, pericardial fat volume, aorta and pulmonary trunk diameters, liver density, aortic calcification, aortic valve calcification, and/or mitral valve calcium.
Using the Agatston score, two individuals can have a CAC 200 score. The individual with a higher number of plaques, or with the lower density of plaques (lower HU), is at higher risk. If there is only one plaque, the individual with the plaque located proximally close to the origin of coronaries is at higher risk than the individual with the plaque located mid or distal. Therefore, the usefulness of a CAC score can be increased if additional information regarding the number, density, distribution, and location of plaques is obtained.
A chest CT scan can be a suitable tool for evaluating the risk of an individual for two leading causes of mortality and morbidity, namely CVD and lung cancer. A chest CT scan can provide information about both the cardiovascular system and the pulmonary system. Applying artificial intelligence (AI) to chest CT scans can empower a provider to rapidly extract actionable information for evaluating the risk of adverse health events not only related to CVD and lungs (that is, detecting pre-cancerous lung nodules, silent emphysema, coronary calcification, aortic calcification, cardiac valves calcification, cardiac chambers enlargement, aortic enlargement, pulmonary artery enlargement, etc.), but it can also detect osteoporosis and osteopenia by measuring thoracic vertebral bone density, cardiometabolic associated excess intrathoracic visceral fat, fatty liver disease, muscle loss or sarcopenia, and other abnormalities in other organs such as thyroid nodules and esophageal masses. All these findings can be actionable and can be automatically detected by a suitably trained artificial intelligence system in both non-contrast and contrast enhanced chest CT scans.
Artificial intelligence (AI) can be used in medical imaging, and AI performance can be fine-tuned to facilitate detection and quantification of various clinical conditions. BMD measurement often requires additional manual work. The systems and methods disclosed here can be a major add-on to CAC and lung cancer scans. The methods and systems disclosed here can be advantageous in identifying individuals at risk of adverse health outcomes even when the individuals are asymptomatic.
CAC score and lung cancer screening CT scans are obtained each year in large quantities. CT-based assessment of BMD measurement for osteoporosis screening, initially manually and then with automated approaches. A fully automated Deep Learning (DL) model has the potential of increasing the reach of CT-based osteoporosis screening.
AI-based detection can measure BMD for both full chest and cardiac CT scans with an accuracy level comparable to manual measurements. For the same patient, there are comparable results obtained between three of the vertebral bones found in cardiac and lung CT scans, respectively. BMD assessment can be carried out with either a cardiac or a lung CT scan and provide the same results clinically for the same patient. In one embodiment, a BMD measurement tool can offer unique and significant advantages, being that patients undergo no additional radiation nor additional trips to imaging centers and there is no additional scanner cost or additional scanning time.
BMD can be calculated from normal bone tissue of trabecular bone in specific spines. Thus, it is important to detect individual vertebrae and identify disk locations for removal from calculation.
Two DL models were developed for cardiac and full chest CT scan images, respectively. The models were trained to detect trabecular bone and disks with a training data set of 132 cardiac and 37 lung CT images. For ground truth, 225 cardiac and lung CT images for whole spine and disk locations were manually segmented. Each DL model has two steps to automatically detect individual vertebrae and disks. In the first step, the models were trained to focus on the whole spine area and trained for 100 epochs. Transfer learning was then used to train for disk locations using the pretrained models. In one embodiment, the architecture of the models can include an encoder and a decoder. The encoder can be a UNet with 12 layers of 2D convolutions, skip connections, Leaky ReLu activations and/or batch normalization. The decoder can be a 3-layer convolution 2D with Leaky ReLu activations and/or a sigmoid at the end.
After identifying individual vertebrae, a set of rules can be used to find the disk location. Vertebrae smaller or bigger than a standard size were removed from calculations. Signal processing was used to erode the borders to segment the entire vertebral bone.
BMD is measured in the trabecular bone in the center of each vertebra. To measure the BMD, the center of mass for each vertebra can be found. A cylinder with a radius of 1-ccm can be created where the center of the mass can be identified. The height of cylinder can be determined by taking two slices above and below adjacent disk locations to avoid cortical bones. The mean HU can be calculated from this volume.
A minimum of three individual vertebrae HU values can be calculated this way. In one embodiment, a case is dismissed if three separate measurements are not found. For each measurement a normalization factor for scanner type can be used to calculate BMD value. Using three vertebrae, T and Z values are calculated considering gender and age of the patient. The input to the model can be cardiac gated CT scans and the output can be, for example, a JSON file that includes the mean HUs, BMD, T score and Z score values. Two snapshots of sagittal and coronal views with superimposed spine and disk segmentations can also be rendered for quality control.
Referencing
At step 115, in one embodiment, an AI analysis from the AIS is received. In certain embodiments, the AI analysis can include, for example, a coronary artery calcium score based, at least in part, on an Agatston score; a measurement of plaque density; a measurement of plaque quantity; one or more locations of plaque; a volume of at least one cardiac structure; and a measurement of cardiac ejection fraction (EF). In one embodiment, the AI analysis is based at least in part on the set of non-contrast CT scan images. In some embodiment, the AI analysis can include one or more measurements of: pericardial fat, thoracic aortic calcification, and bone mineral density.
At step 120, based at least in part on the AI analysis, a patient's risk for an adverse health outcome is determined. In one embodiment, at least one estimate of risk for an adverse health outcome is determined, at least in part, by a computer-enabled risk calculator. For example, based at least on a volume of at least one cardiac structure, a computer-enabled risk calculator can compute a patient's risk for heart failure. By way of another example, based at least on EF, a computer-enabled risk calculator can compute a risk of future heart failure, myocardial infarction, and/or sudden cardiac death. In certain embodiments, the adverse health outcome can be osteopenia and/or osteoporosis.
In certain embodiments, system 200 can include CT scan images 210, which can be stored in a computer memory, for example. CT scan images can include images obtained from CT scans, for example, contrast enhanced CT scans, non-contrast enhanced CT scans, ECG-gated cardiac CT scans, non-gated cardiac CT scans, non-gated full chest CT scans, low dose lung cancer screening CT scans, and a combination of contrast enhanced and non-contrast enhanced chest CT scans. In some embodiments the cardiovascular structure can be, for example, left atrium (LA), left ventricle (LV), left ventricular wall (LVW), right atrium (RA), right ventricle (RV), aorta, and/or pulmonary artery.
System 200 can, in one embodiment, include risk calculator 215 configured to take as input one or more of the measurements or scores from the AI-based analysis to determine, based at least in part on the measurements or scores, a risk of a patient for an adverse health condition. In certain embodiments, the adverse health condition can be, for example, atrial fibrillation (AF), heart failure (HF), stroke, cerebrovascular events, chronic obstructive pulmonary diseases (COPD), emphysema, ischemic heart disease, cardiovascular mortality, and all-cause mortality. In some embodiments, the risk is determined based on the estimated measurement and/or score and considering other variables, such as patient's age, gender, height, weight, body surface area, body mass index, and ethnicity, for example. In some embodiments, the estimated volume can be used with one or more health related variables to enhance the prediction model, resulting in a multivariate composite index of health to better determine the risk of future adverse health conditions. In certain embodiments, the one or more health related variables can be, for example, blood pressure, heart rate, blood oxygenation, blood tests, medications, and other patient medical data. In some embodiments, AI-based analyzer 205 and risk calculator 215 can be implemented as a single component that, for example, receives CT scan images 210 as input and returns risk estimates as output shown in display 220. In certain embodiments, risk calculator 215 can be a comprehensive risk assessment module configured to combine the output of AIS2 with, for example, known emerging risk factors. Said risk assessment module can be, for example, an AIS classifier or a multi-variate model that takes into account CT scan analysis and non-image data.
Polygenic risk score (PRS) is a quantitative measure that combines information from numerous genetic markers across an individual's genome to estimate their genetic predisposition for developing CVD. In some embodiments, risk calculator 215 can use PRS to provide valuable insights into an individual's susceptibility to CVD, thereby enabling early identification of high-risk individuals based on a comprehensive set of data that includes genetic susceptibility information, imaging biomarkers, epidemiological risk factors, circulating biomarkers, and physiological information. Use of PRS in system 200 facilitates improving risk assessment for early preventive strategies, and personalized treatment approaches for CVD. PRS is primarily a long-term indicator of risk; however, environment-genes interactions make the role of PRS a lot more complex. In certain embodiments of system 200, an AIS can be configured to stratify risk based on the weight of each component per individual. For example, the role of PRS in an 80-year-old male can be assigned less weight than the physiological findings or imaging information from the heart. In contrast, in a 40-year-old male the PRS can be assigned a higher weight and, consequently, plays a greater role in developing a personalized risk reduction and preventive plans.
In one embodiment, risk calculator 215 can include a software module configured to determine, for example, the ten-year risk of Hard CVD outcome for a female individual by implementing the formula: 10 yr risk=1−(0.9757{circumflex over ( )}exp(0.0635*age −0.0190*HDL+0.4353*(HTN)+0.8623*Smoker+0.0400*LAVI+0.0518*LVVI −0.0566*RVVI −3.3674)). Similar equations can be derived and implemented based on exemplary values provided in the tables below. It should be noted that not all variables need be included in a given risk calculator formula.
Here, age is the individual's age, total cholesterol is the individual's total cholesterol measurement, HDL is the individual's HDL measurement, HTN is presence of hypertension, smoker is a binary variable indicating status as smoker or non-smoker, LAVI is the individual's left atrium volume index, LVVI is the individual's left ventricle volume index, and RVVI is the individual's right ventricle volume index. In some embodiments, LAVI, LVVI, and RVVI can be determined by dividing the cardiac chamber volume by the individual's weight. This results in normalizing the cardiac chamber volume and making it comparable across individuals.
In certain embodiments, system 200 can include display 220 configured for displaying, for example, cardiovascular structures, estimated volumes, and/or a graphic representation of the risks determined by risk calculator 215. In one embodiment, computer enabled display 220 can be on a mobile application or a web application that can be used by patients and/or care providers. In some embodiments, computer enabled display 220 can be on a desktop application run on premises to, for example, avoid patient data security concerns.
CARDIAC EJECTION FRACTION MEASUREMENT USING NON-CONTRAST CT SCANS. Cardiac ejection fraction also referred to as left ventricular ejection fraction (EF) is a measurement of the amount of blood pumped out of the heart into the body through the aorta with each heartbeat. EF is typically expressed as a percentage, and it can be an important indicator of heart health. A common method of measuring EF uses echocardiography, which uses ultrasound waves to create images of the heart. The images can be used to measure the blood volume in the heart before a heartbeat (EDV) and after a heartbeat (ESV). EF is calculated by dividing the blood volume pumped out of the heart (EDV-ESV) by EDV.
Other methods for measuring EF use magnetic resonance imaging (MRI) and computed tomography (CT) scans. These imaging techniques can provide more detailed images of the heart, but they are more expensive and may not be readily available in all healthcare settings. Gated-ECG, contrast-enhanced CT scans can be used to measure EF, and such scans can be useful in certain clinical situations where other methods are not feasible.
Contrast agents are substances that are injected into the body to make tissues or structures more visible for imaging. However, injecting a patient with a contrast agent can damage, for example, the kidneys. Contrast-induced nephropathy (CIN) is a type of kidney damage that can occur after a patient receives a contrast agent for medical imaging procedures, such as CT scans, angiograms, or X-rays. In severe cases, CIN can lead to acute kidney injury or even kidney failure, requiring dialysis or other forms of treatment. The risk of developing CIN depends on several factors, including the type and dose of contrast agent used, the patient's age, underlying medical conditions such as kidney disease or diabetes, and the presence of other risk factors such as dehydration or the use of certain medications.
During CT scans, a patient's heart rate and rhythm can be monitored using an electrocardiogram (ECG). ECG is a non-invasive medical test that records the electrical activity of the heart over a period of time. The ECG signals can be used to synchronize the CT scanner with the patient's heartbeat, allowing acquisition of heart images at specific points in the cardiac cycle.
Retrospective ECG gating and prospective ECG gating are two methods used in CT imaging to reduce the effect of motion artifacts from the beating heart during image acquisition. The main difference between these two methods is the timing of image acquisition during the cardiac cycle.
Retrospective ECG gating involves acquiring ECG data and images continuously and simultaneously throughout the cardiac cycle. This method uses a computer to sort the ECG data into separate phases of the cardiac cycle, typically using the R-wave of the ECG signal as a reference point. The images acquired are then sorted into different phases of the cardiac cycle, and reconstructed into a final image that represents a single phase of the cycle. Retrospective ECG gating provides high temporal resolution and allows for the reconstruction of images at any point in the cardiac cycle. However, compared to prospective ECG gating, retrospective ECG gating exposes the patient to a much higher level of X-ray radiation.
Retrospective ECG gating is generally used to measure cardiac function. Because images are acquired throughout the cardiac cycle, volume measurements of the right and left ventricles can be obtained in the end-systole phase (ES) and end-diastole phase (ED), allowing the calculation of stroke volume, EF, and cardiac output. Retrospective ECG gating can also facilitate acquiring diagnostic images of the coronary arteries. In contrast to prospective ECG gating, retrospective ECG gating facilitates ECG editing to remove artifacts related to premature ventricular contractions and/or dropped beats.
Prospective ECG gating involves acquiring images only during a specific phase of the cardiac cycle, typically during diastole when the heart is at rest. The ECG signal is used to trigger image acquisition only during this phase of the cycle. This method provides lower temporal resolution than retrospective ECG gating, but the advantage is that it reduces radiation exposure.
An AI-enabled technique can be used for automated cardiac chambers volumetry (for example, AutoChamber) in non-contrast CT scans (for example, chest CT scans). The technique estimates cardiac chamber volume (including, for example, LV volume) using non-contrast, gated and/or non-gated CT scans. This technique can detect individuals (even asymptomatic individuals) at high risk of adverse health conditions such as heart failure, atrial fibrillation, and stroke.
In one embodiment, the invention is directed to a method of measuring EF. The method can include acquiring ECG-gated CT scans during ED (usually around 75%) and at ES during “isovolumic relaxation” which is around 25%. In some embodiments, using AutoChamber, the change in LV volume is measured to estimate EF. The change in the entire LV and LV wall can also be measured to estimate EF. Compared to retrospective ECG-gating, which requires contrast-enhanced CT scans, this AI-enabled technique reduces X-ray radiation significantly and does not require injecting patients with contrast agents.
In certain embodiments, this technique can be used in addition to screening with AutoChamber. For example, if there is an indication of an enlarged chamber, then a second scan can be done to measure EF. The second scan can be done immediately after the first scan and can be limited to ES only, or it can be a full-set of ED plus ES snapshot images. In the former scenario, AutoChamber can be embedded in the CT scan to facilitate an immediate decision on whether to obtain an EF measurement.
In one embodiment, the accuracy of the EF measurement can be improved by acquiring images during the “isovolumic contraction” phase and the “isovolumic relaxation” phase. During isovolumic contraction and isovolumic relaxation the volume is, respectively, maximum and minimum. In some embodiments, this method can be used as an embedded, AI-enabled algorithm of any ECG-gated, CT scan used for cardiac imaging.
At steps 325 and 330, measurements for end-diastole left ventricle (LV) volume (“first measurement”) and end-systole LV volume (“second measurement”) are received. In one embodiment, said measurements are based, at least in part, on the end-diastole and end-systole images (335). At step 340, an EF measurement is determined based, at least in part, on the first and second measurements.
In one embodiment, the first and second images are acquired during a prospective ECG-gated CT scan. In some embodiments, the first image is acquired during isovolumic contraction, and the second image is acquired during isovolumic relaxation. In certain embodiments, a measurement of LV wall volume can be received from the AIS.
Heart failure with preserved ejection fraction (HFpEF) is typically defined as heart failure with a left ventricular ejection fraction (LVEF) of 50% or greater. Heart failure with a reduced ejection fraction (HFrEF) is heart failure with an LVEF of 40% or less.
The table below shows the improvements of Agatston+ over prior art Agatston CAC score in prediction of various adverse health outcomes at 5-year, 10-year, and 15-year follow up periods. This table shows the net reclassification index (NRI) of cardiac chambers volumetry when added to the Agatston CAC score. Any NRI over 0.1 is viewed as clinically meaningful and NRI over 50 is viewed as very effective.
In certain embodiments, a modified Agatston Score can be produced by taking into account one or more of the following: 1) the location of plaque (whether the plaque is in the beginning, middle, or end of a coronary artery); 2) the number of coronary arteries that have a plaque; 3) the mean density of HU units in a plaque (in the prior art Agatston score only the maximum intensity is counted); 4) the standard deviation of HU per plaque; and 5) the distance between two plaques or multiple plaques. In some embodiments, if there is low or no calcium, the process is performed again with HU 100-130 (which is 30 units below the prior art Agatston score cut-off point) to look for tiny amount of calcium. Using an AIS trained on background noises in non-contrast CT scan and on 3D coronary artery zones, this method can facilitate detecting coronary arteries with non-calcified (that is, soft plaques) and raise flags for further medical attention including a recommendation of contrast enhanced coronary CT angiography (CCTA) to provide delineation of coronary artery wall.
In certain embodiments of this invention, and referencing here U.S. provisional application 63/414,561, obtaining a CCTA after a non-contrast CT can facilitate identifying plaques with excessive vasa vasorum (angiogenesis) surrounding the plaque that is active or inflamed. The more vasa vasorum the more blood flow and contrast agent surrounding the plaque therefore showing higher HU. An active plaque is usually less calcified than a passive or stable plaque. By detecting passive plaques with CAC score and detecting active plaques with vasa vasorum density, embodiments of this invention can cover significant aspects of coronary risk assessment.
One motivation for the inventive embodiments described herein is the need for systems and methods to facilitate a more accurate prediction of adverse health outcomes, as compared to the current state of the art. Embodiments of the invention, for example with a combination of cardiac chambers volumetry and a coronary calcium score, can provide an unprecedented ROC AUC of 0.90 for prediction of stroke in 12 months. With the prior art Agatston score, about 3% of individuals with Agatston Score CAC=0 in a 15-yr follow up study had a heart attack. With the modified Agatston score (“Agatston+”) produced by the systems and methods disclosed here, it can facilitate assigning a more accurate and reliable CAC score of zero.
This application claims priority to U.S. patent applications 63/414,546 and 63/414,561, both filed on Oct. 9, 2022, and are hereby incorporated herein in their entireties by reference. This application is a continuation-in-part of co-pending U.S. patent application Ser. No. 18/167,691, filed on Feb. 10, 2023, which is hereby incorporated herein in its entirety by reference. This application is a continuation-in-part of co-pending U.S. patent application Ser. No. 18/167,852, filed on Feb. 11, 2023, which is a continuation-in-part of U.S. application Ser. No. 17/657,754, filed on Apr. 2, 2022, which are both hereby incorporated herein in their entireties by reference.
| Number | Date | Country | |
|---|---|---|---|
| 63414546 | Oct 2022 | US | |
| 63414561 | Oct 2022 | US |
| Number | Date | Country | |
|---|---|---|---|
| Parent | 18167852 | Feb 2023 | US |
| Child | 18356444 | US | |
| Parent | 18167691 | Feb 2023 | US |
| Child | 18167852 | US |
