Radiomic signature of epicardial regions

Abstract

A method for characterizing an epicardial region using medical imaging data (e.g., a gray level co-occurrence matrix) of a subject. The method includes calculating a value of an epicardial radiomics signature for the epicardial region using the medical imaging data. A method for deriving an epicardial radiomics signature indicative of cardiac health is also disclosed. The method includes constructing an epicardial radiomics signature using a radiomics dataset. A system for performing the foregoing method is also disclosed.

Description

Technical Field

[0001] The present invention relates to a method for characterizing epicardial regions, in particular using radiomic signatures, and a system for use in said method. The present invention also relates to a method for deriving such signatures and a system for use in said method. Background Art

[0002] Cardiac (i.e., heart) pathologies, including ischemic heart disease, heart failure, cardiomyopathy, and arrhythmias such as atrial fibrillation (AF), represent a global burden on healthcare systems and patients' quality of life, as does ischemic stroke. Although it is feasible to study cardiac function through imaging (e.g., by echocardiography or cardiac magnetic resonance), there is currently no way to study myocardial tissue phenotype and disease (e.g., atrial redox state, fibrosis, or pro-fibrotic signaling), which is often the underlying cause of such pathologies.

[0003] Epicardial obesity is an independent risk factor for cardiac disease and stroke. For example, in clinical studies using imaging, expansion of epicardial adipose tissue volume was found to be independently associated with the risk of atrial fibrillation (Antonopoulos, AS and Antoniades, C., The role of epicardial adipose tissue in cardiac biology: classic concepts and emerging roles. J Physiol doi: 10.1113 / JP273049 (2017)). Vice versa, atrial disease leads to fatty infiltration of the atrial myocardium by enhancing natriuretic peptide signaling-induced adipogenesis (Suffee, N. et al., Atrial natriuretic peptide regulates adipose tissue accumulation in adult atria. Proc Natl Acad Sci U SA 114, E771-E780, doi: 10.1073 / pnas.1610968114 (2017)).

[0004] Computed tomography is now increasingly used for imaging of human adipose tissue (e.g., quantifying fat volume), and volumetric assessment of epicardial obesity has been successfully used for cardiac risk assessment in clinical studies (Antonopoulos, AS et al., Mutual Regulation of Epicardial Adipose Tissue and Myocardial Redox State by PPAR-gamma / Adiponectin Signalling, Circ Res 118, 842-855, doi: 10.1161 / CIRCRESAHA.115.307856 (2016)). Specifically, CT imaging has been used as a noninvasive modality to study human obesity by identifying voxels of fat based on their CT attenuation or radiodensity (e.g., -190 to -30 Hounsfield units), and fat volume measurements have been established as prognostic biomarkers for cardiovascular disease risk (Rosito, GA et al., Pericardial fat, visceral abdominal fat, cardiovascular disease risk factors, and vascular calcification in a community-based sample: the Framingham Heart Study. Circulation 117, 605-613, doi: 10.1161 / CIRCULATIONAHA.107.743062 (2008)).

[0005] Although this volumetric approach can be used to reliably describe and quantify epicardial obesity, it is only indirectly and weakly linked to features of adipose tissue biology and underlying myocardial biology and disease. Other methods that rely on direct assessment of cardiac (e.g., atrial) function or volume (e.g., echocardiography or cardiac magnetic resonance) cannot provide reliable information about cardiac tissue phenotype and disease. Therefore, there is an unmet need for noninvasive methods for assessing cardiac health, specifically myocardial health, and for detecting or diagnosing myocardial diseases such as myocardial oxidative stress, inflammation and / or fibrosis. There is also a need for noninvasive methods to assess the risk of developing cardiac conditions, specifically the risk of developing arrhythmias such as AF. There is also a need for noninvasive methods for assessing the risk of ischemic stroke. Summary of the invention

[0006] According to a first aspect of the invention, a method for characterizing an epicardial region (e.g., its phenotype, e.g., composition and / or texture) using medical imaging data of a subject is provided. The method can be used to indirectly characterize cardiac tissue, e.g., myocardium. Specifically, the method can be used to characterize or assess myocardial health of myocardium that is specifically adjacent to the epicardial region. For example, the method can be used to detect or identify myocardial diseases, such as fibrosis, inflammation and / or oxidative stress, specifically fibrosis. The method can include using medical imaging data to calculate a value of an epicardial radiomics signature of an epicardial region. The epicardial radiomics signature can be calculated using measured values ​​of a plurality of epicardial radiomics features of the epicardial region or based on the measured values. The measured values ​​of the epicardial radiomics features can be calculated based on or by the medical imaging data.

[0007] The epicardial radiomics signature may provide a measure of texture of the epicardial region. At least one of the epicardial radiomics features may provide a measure of texture of the epicardial region, for example, at least one of the epicardial radiomics features may be a texture statistic.

[0008] The epicardial radiomic signature (i.e., its value) can be indicative of cardiac health, specifically myocardial health. For example, the epicardial radiomic signature can be indicative of or associated with (e.g., statistically significantly associated with) myocardial disease. The epicardial radiomic signature can be indicative of myocardial inflammation. The epicardial radiomic signature can be indicative of myocardial redox state or oxidative stress. The epicardial radiomic signature can be indicative of myocardial fibrosis.

[0009] The epicardial radiomic signature (i.e., its value) can predict the likelihood that a subject suffers from a cardiac or heart condition. A cardiac condition can be associated with a myocardial disease, specifically associated with myocardial fibrosis. A cardiac condition can be an arrhythmia (e.g., atrial fibrillation), ischemic heart disease, heart failure, and / or cardiomyopathy.

[0010] The epicardial radiomic signature can predict the likelihood that a subject will experience or suffer a stroke, particularly an ischemic stroke. The epicardial radiomic signature can be associated with or indicative of a risk of stroke, e.g., an increased or high risk of stroke compared to the general population.

[0011] The epicardial region may include or consist of a peri-atrial region, for example, a peri-left atrial region. The peri-left atrial region may include or consist of an epicardial region adjacent to the atrial septum and / or the anterior wall of the left atrium (i.e., the region immediately in front of the left atrium). These regions include or consist of epicardial adipose tissue and epicardial connective tissue.

[0012] At least one of the epicardial radiomics features may be calculated based on a wavelet transform of the attenuation values.

[0013] As identified in Table 3, the plurality of epicardial radiomic features may include at least two epicardial radiomic features selected from groups 1 to 15 of epicardial radiomic features. The at least two epicardial radiomic features may each be selected from a different group. Groups 1 to 15 may be limited to those epicardial radiomic features that have a correlation with important epicardial radiomic features of |rho|≥0.800. Groups 1 to 15 may be limited to those epicardial radiomic features that have a correlation with important epicardial radiomic features of |rho|≥0.850. Groups 1 to 15 may be limited to those epicardial radiomic features that have a correlation with important epicardial radiomic features of |rho|≥0.900. Groups 1 to 15 may be limited to those epicardial radiomic features that have a correlation with important epicardial radiomic features of |rho|≥0.950.

[0014] As identified in Table 3b, the plurality of epicardial radiomic features may include at least two epicardial radiomic features selected from groups 1 to 16 of epicardial radiomic features. The at least two epicardial radiomic features may each be selected from a different group. Groups 1 to 16 may be limited to those epicardial radiomic features that have a correlation with important epicardial radiomic features of |rho|≥0.800. Groups 1 to 16 may be limited to those epicardial radiomic features that have a correlation with important epicardial radiomic features of |rho|≥0.850. Groups 1 to 16 may be limited to those epicardial radiomic features that have a correlation with important epicardial radiomic features of |rho|≥0.900. Groups 1 to 16 may be limited to those epicardial radiomic features that have a correlation with important epicardial radiomic features of |rho|≥0.950.

[0015] At least two epicardial radiomics features may include at least two of the following: inverse moment HHH, minimum value LHH, low grayscale area advantage LLL, maximum probability LLL, busyness LHH, regional entropy LLL, run entropy LLL, maximum 3D diameter, long run high grayscale advantage LLL, joint average LLL, difference entropy LLL, sum entropy HHH, difference entropy LHH, sum of squares LHH and grayscale variance LHH (GLDM).

[0016] At least two epicardial radiomics features can be composed of 15 epicardial radiomics features and can be composed of the following: inverse difference moment HHH, minimum value LHH, low grayscale area advantage LLL, maximum probability LLL, busyness LHH, regional entropy LLL, run entropy LLL, maximum 3D diameter, long run high grayscale advantage LLL, joint average LLL, difference entropy LLL, sum entropy HHH, difference entropy LHH, sum of squares LHH and grayscale variance LHH (GLDM).

[0017] The at least two epicardial radiomics features may include at least two of the following: 10th percentile, IDHHL, variance, normalized grayscale nonuniformity (GLRLM), dependent variance HLL, size regional nonuniformity LHL, skewness, root mean square, grayscale nonuniformity LLH, large area predominance LLH, IDMN HHH, area percentage HHL, kurtosis, normalized size regional nonuniformity HHH, difference entropy LLL, and autocorrelation HHL.

[0018] At least two epicardial radiomic features can be composed of 16 epicardial radiomic features and can be composed of the following: 10th percentile, ID HHL, variance, normalized grayscale unevenness (GLRLM), dependent variance HLL, size regional unevenness LHL, skewness, root mean square, grayscale unevenness LLH, large area dominance LLH, IDMN HHH, area percentage HHL, kurtosis, normalized size regional unevenness HHH, difference entropy LLL, and autocorrelation HHL.

[0019] As identified in Table 1, Table 2, or Table 3, the plurality of epicardial radiomic features may include at least two epicardial radiomic features selected from epicardial radiomic features of clusters A through D. The at least two epicardial radiomic features may each be selected from a different cluster.

[0020] As identified in Table 1b, Table 2b, or Table 3b, the plurality of epicardial radiomic features may include at least two epicardial radiomic features selected from epicardial radiomic features of clusters A through D. The at least two epicardial radiomic features may each be selected from a different cluster.

[0021] The plurality of epicardial radiomic features may include at least two epicardial radiomic features selected from epicardial radiomic features of clusters A to D, wherein: cluster A is composed of inverse moment HHH, minimum value LHH, regional variance LLL, low grayscale regional advantage LLL, maximum probability LLL, elongation, cluster shadow LLL, busyness LHH, grayscale unevenness LLL and skewness HHH; cluster B is composed of regional entropy LLL, cluster prominence LLL, grayscale variance LLL (GLDM) and run entropy L Cluster C is composed of the minimum axis, the maximum 2D diameter row, the major axis, the maximum 2D diameter column, the maximum 2D diameter slice and the maximum 3D diameter; and cluster D is composed of autocorrelation LLH, long run high grayscale advantage LLL, joint average LLL, autocorrelation LLL, difference entropy LLL, difference entropy LLH, sum of squares HLH, sum of squares HHH, sum entropy HHH, grayscale variance LLH (GLDM), difference entropy LHH, sum of squares LHH and grayscale variance LHH (GLDM). At least two epicardial radiomics features can be selected from different clusters.

[0022] The plurality of epicardial radiomic features may include at least two epicardial radiomic features selected from epicardial radiomic features of clusters A to D, wherein: cluster A is composed of inverse moment HHH, minimum value LHH, low grayscale area advantage LLL, maximum probability LLL, and busyness LHH; cluster B is composed of regional entropy LLL and run entropy LLL; cluster C is composed of maximum 3D diameter; and cluster D is composed of long run high grayscale advantage LLL, joint average LLL, difference entropy LLL, sum entropy HHH, difference entropy LHH, square sum LHH, and grayscale variance LHH (GLDM). The at least two epicardial radiomic features may each be selected from a different cluster.

[0023] The at least two epicardial radiomic features may include at least three epicardial radiomic features. The at least two epicardial radiomic features may include at least four epicardial radiomic features. The at least two epicardial radiomic features may include at least 15 radiomic epicardial features. The at least two epicardial radiomic features may include at least 16 epicardial radiomic features.

[0024] The medical imaging data may include an attenuation value for each voxel of a plurality of voxels corresponding to at least the epicardial region.

[0025] The method may further include identifying an epicardial region using the medical imaging data. The epicardial region is identified using a manual outline. The epicardial region may contain only voxels of the medical imaging data whose attenuation values ​​fall within a given range of attenuation values. The given range may be about -190 to about -30 Hounsfield units.

[0026] The method may further include identifying the epicardial region using an automatic segmentation algorithm.The segmentation algorithm may be trained using machine learning, in particular deep learning, to segment the medical imaging data.

[0027] The epicardial region may include or consist of only voxels of the medical imaging data, the attenuation values ​​of which fall within a given range of attenuation values. The given range of attenuation values ​​may correspond to fat and / or connective tissue.

[0028] A given range may include or encompass attenuation values ​​above and below -30 Hounsfield units. A given range may be about -190 to about -30 Hounsfield units or about -190 to about +150 Hounsfield units. Alternatively, a given range may include or encompass these ranges.

[0029] The method may further include segmenting the epicardial region. Values ​​of a plurality of epicardial radiomic features may be calculated based on the segmented epicardial region.

[0030] The value of each of the plurality of epicardial radiomics features may be calculated based on raw attenuation values, binned attenuation values, or a wavelet transform of attenuation values.

[0031] The method may further include predicting the risk of the subject suffering from a cardiac condition based at least on the calculated value of the epicardial radiomic signature. The cardiac condition may be an arrhythmia, such as atrial fibrillation. For example, the method may further include predicting the risk of the subject suffering from a post-operative arrhythmia.

[0032] The method may further include predicting or classifying the risk of the subject experiencing or suffering an ischemic stroke based at least on the calculated value of the epicardial radiomic signature.

[0033] The method may further include identifying whether the individual is at risk for stroke, eg, high or increased risk for stroke, eg, relative to the general population, based on the calculated value of the epicardial radiomics signature.

[0034] The method may further include administering or prescribing a preventive treatment, such as a medication, to the individual to reduce the risk of stroke if the individual is identified as being at risk for stroke.

[0035] The method may further include assessing cardiac health or myocardial health of the subject based at least on the calculated value of the epicardial radiomic signature.

[0036] The method may further include determining whether the subject suffers from a myocardial disease based at least on the calculated value of the epicardial radiomic signature. The myocardial disease may be fibrosis. The myocardial disease may be inflammation. The myocardial disease may be oxidative stress.

[0037] The epicardial radiomics signature may be calculated using, based on (ie, may consist of or include) a decision tree, in particular a regression tree.

[0038] The epicardial radiomics signature may include a weighted sum of multiple epicardial radiomics features. The epicardial radiomics signature may be linearly related to the weighted sum of multiple epicardial radiomics features.

[0039] According to a second aspect of the present invention, a method for deriving an epicardial radiomics signature is provided. The radiomics signature may indicate cardiac health or disease, specifically myocardial health or disease. The radiomics signature may indicate or predict the risk of stroke. For example, the radiomics signature may be suitable for identifying individuals at risk of stroke. The epicardial radiomics signature may be suitable for indirectly characterizing cardiac tissue, for example, myocardium. Specifically, the signature may be suitable for characterizing or evaluating or may indicate myocardial health or disease, specifically myocardium adjacent to an epicardial region. For example, the signature may indicate or may be suitable for detecting or identifying myocardial diseases, such as fibrosis, inflammation and / or oxidative stress, specifically fibrosis. The method may include using a radiomics dataset to construct an epicardial radiomics signature indicating cardiac health or disease. The epicardial radiomics signature may be calculated based on a second plurality of epicardial radiomics features of an epicardial region. The dataset may include values ​​of a first plurality of epicardial radiomics features of an epicardial region, the values ​​of the first plurality of epicardial radiomics features being obtained from medical imaging data of the epicardial region for each of a plurality of individuals. The plurality of individuals may include a first group of individuals having or identified as having (at the time of collecting the medical imaging data) or previously having a cardiac condition or myocardial disease or a history of stroke and a second group of individuals not having or identified as not having (at the time of collecting the medical imaging data) a cardiac condition or myocardial disease or a history of stroke, and optionally not having a history of a cardiac condition. For example, the second plurality of epicardial radiomic features are selected from the first plurality of epicardial radiomic features based on an analysis of a data set or using the data set. Specifically, the second plurality of epicardial radiomic features may be selected from the first plurality of epicardial radiomic features to provide an epicardial radiomic signature indicative of cardiac health and / or disease and / or indicative of a risk of developing a cardiac condition or experiencing a stroke, as determined based on the data set, for example, using a machine learning algorithm.

[0040] The method may further include using the data set to identify important epicardial radiomic features from a first plurality of epicardial radiomic features, each of the first plurality of epicardial radiomic features being identified as being significantly associated with a cardiac condition or myocardial disease or a history of stroke, as determined according to the data set. A second plurality of epicardial radiomic features may include at least two epicardial radiomic features, the at least two epicardial radiomic features being selected as different important epicardial radiomic features or being co-linear therewith (specifically identified as being co-linear therewith). Each epicardial radiomic feature of at least two of the second plurality of epicardial radiomic features may be selected as being significantly associated with a cardiac condition or myocardial disease or a history of stroke, as determined according to the data set.

[0041] The method may further include using a feature selection algorithm (e.g., a machine learning feature selection algorithm) to identify a subset of epicardial radiomic features (optionally, a subset of important epicardial radiomic features) based on which the radiomic signature should be calculated (i.e., the best features). In other words, the subset of predicted radiomic features can maximize the accuracy of the epicardial radiomic signature (e.g., optimize or maximize the association of the radiomic signature with cardiac symptoms or myocardial disease or a history of stroke), for example, when the epicardial radiomic signature is calculated (only) based on the subset of epicardial radiomic features (and optimized). The subset of epicardial radiomic features can maximize the association of the preliminary epicardial radiomic signature with cardiac symptoms or myocardial disease or a history of stroke (when the preliminary epicardial radiomic signature is calculated based on a subset of important epicardial radiomic features), as determined according to the data set. At least two epicardial radiomic features may be selected as different epicardial radiomic features belonging to a subset, or collinear with the different epicardial radiomic features of the subset. The at least two epicardial radiomic features may include all epicardial radiomic features belonging to the subset, or their collinear equivalents that are collinear with the epicardial radiomic features, as determined based on the data set. In other words, the at least two epicardial radiomic features may include each epicardial radiomic feature of the epicardial radiomic features belonging to the subset, or an epicardial radiomic feature that is collinear with each epicardial radiomic feature of the epicardial radiomic features.

[0042] The method may further include identifying groups of epicardial radiomic features, each of the groups including a significant epicardial radiomic feature and one of its collinear equivalents that is collinear with the significant epicardial radiomic feature, as determined from the data set.At least two epicardial radiomic features may be selected from different groups.

[0043] The method may further include identifying multiple clusters of important epicardial radiomic features by performing a cluster analysis (e.g., using a clustering algorithm, specifically a machine learning clustering algorithm), for example, a correlation cluster analysis. At least two epicardial radiomic features can each be selected from a different cluster or selected to be co-linear with important epicardial radiomic features from different clusters. The cluster analysis can identify clusters based on the strength of correlations between important epicardial radiomic features. Intra-cluster correlations can be stronger than inter-cluster correlations. For example, the epicardial radiomic features within each cluster can be more correlated with each other than with epicardial radiomic features in other clusters.

[0044] The cluster analysis can be a hierarchical cluster analysis, a k-means cluster analysis, a distribution-based cluster analysis, or a density-based cluster analysis. Specifically, the cluster analysis can be a hierarchical cluster analysis. The cluster analysis, e.g., a clustering algorithm, can identify clusters based on distances between epicardial radiomics features, e.g., squared Euclidean distances between epicardial radiomics features, for example, in a correlation plot. The clustering algorithm can identify clusters based on distances between features in a correlation space, where the distance between each pair of features corresponds to the degree to which those features are correlated, i.e., the closer two features are in the correlation space, the more correlated the two features are to each other.

[0045] If two epicardial radiomic features have a degree of correlation equal to at least a correlation threshold, they can be identified as collinear. The Spearman rho coefficient can be used to calculate the correlation between epicardial radiomic features. Alternatively, other measures of pairwise correlation, such as the Pearson correlation coefficient (Pearson's r), can be used to calculate collinearity between epicardial radiomic features. The correlation threshold can be at least about |rho|=0.75.

[0046] If the epicardial radiomics feature is associated with a cardiac symptom or myocardial disease or a history of stroke above a significance threshold, it can be determined to be significantly associated with a cardiac symptom or myocardial disease or a history of stroke. The significance threshold can be at least about α=0.05, for example, about α=0.05. A statistical correction for correcting multiple comparisons (such as the Benjamini-Hochberg correction) can be applied to the significance threshold. As will be readily appreciated by those skilled in the art, the association of the epicardial radiomics feature with a cardiac symptom or myocardial disease or a history of stroke can be calculated based on a receiver operating characteristic (ROC) curve analysis, particularly using an area under the curve (AUC) measurement (i.e., C statistic).

[0047] Using the data set or as determined according to the data set, the epicardial radiomics signature can be constructed to be associated with a history of cardiac symptoms or myocardial disease or stroke. Using the data set or as determined according to the data set, the epicardial radiomics signature can be constructed to be significantly associated with a history of cardiac symptoms or myocardial disease or stroke. If the epicardial radiomics signature is associated with a history of cardiac symptoms or myocardial disease or stroke above a significance threshold, it can be determined to be significantly associated with a history of cardiac symptoms or myocardial disease or stroke, as determined according to the data set. The significance threshold can be at least about α=0.05, for example, about α=0.05. As will be readily understood by those skilled in the art, the association of the epicardial radiomics signature with a history of cardiac symptoms or myocardial disease or stroke can be calculated based on a receiver operating characteristic (ROC) curve analysis, particularly using an area under the curve (AUC) measurement (i.e., C statistics).

[0048] The dataset may be divided into training cohort data and validation cohort data. The step of constructing the epicardial radiomics signature may include deriving the signature using the training cohort dataset and validating the signature using the testing cohort dataset.

[0049] The step of constructing the epicardial radiomic signature may include refining the epicardial radiomic signature to increase the association of the epicardial radiomic signature with cardiac symptoms or myocardial disease or a history of stroke.

[0050] The step of constructing an epicardial radiomics signature may be performed using multiple cross validation. The step of constructing an epicardial radiomics signature may be performed using recursive feature elimination, for example, using a random forest algorithm.

[0051] The step of constructing an epicardial radiomics signature can be performed using a machine learning algorithm. For example, the step of constructing an epicardial radiomics signature can include, for example, using a feature selection machine learning algorithm to identify a subset of important epicardial radiomics features to be included in the signature. Feature selection can be performed using recursive feature elimination, for example, using a random forest algorithm. Then, the step of constructing an epicardial radiomics signature can include refining or optimizing the epicardial radiomics signature based on the subset of epicardial radiomics features identified in the feature selection step, for example, using a machine learning algorithm. For example, a second machine learning algorithm (e.g., a radiomics signature optimization algorithm) can be used to, for example, use multiple cross-validation to refine or optimize the epicardial radiomics signature calculated using a subset of epicardial radiomics features. In other words, the second machine learning algorithm refines or optimizes the epicardial radiomics signature calculated based on a subset of important epicardial radiomics features. The second machine learning algorithm can be a decision tree learning algorithm.

[0052] A decision tree can be used to calculate the epicardial radiomics signature. A regression tree can be used to calculate the epicardial radiomics signature.

[0053] The epicardial radiomics signature can be constructed to provide a measure of the texture of the epicardial region.

[0054] At least one of the second plurality of epicardial radiomic features, for example, at least one epicardial radiomic feature of the at least two epicardial radiomic features, can provide a measure of texture of the epicardial region. For example, each of the second plurality of epicardial radiomic features can provide a measure of texture of the epicardial region (i.e., each of the epicardial radiomic features can be a texture statistic).

[0055] A cardiac condition can be associated with (e.g., at least partially caused by or related to) cardiac health or disease, specifically myocardial health or disease. For example, a cardiac condition can be associated with myocardial redox state or oxidative stress. For example, a cardiac condition can be associated with myocardial inflammation. Specifically, a cardiac condition can be associated with myocardial fibrosis. Thus, an epicardial radiomic signature can be constructed to be associated with or indicative of cardiac health or disease (e.g., myocardial fibrosis).

[0056] The myocardial disease may be inflammation, oxidative stress or fibrosis, in particular fibrosis.

[0057] A heart condition can be an arrhythmia. An arrhythmia can be atrial fibrillation.

[0058] The method may further include calculating a value of the derived epicardial radiomic signature for the epicardial region of the patient or subject. For example, the method may further include characterizing the epicardial region of the patient or subject by calculating a value of the derived epicardial radiomic signature. The value of the derived radiomic signature may be calculated based on or using medical imaging data of at least the epicardial region of the patient or subject. The value of the derived radiomic signature may be calculated using or based at least on the values ​​of a second plurality of radiomic features of the epicardial region of the patient or subject.

[0059] Therefore, the method can be used to derive an epicardial radiomic signature and use the derived radiomic signature to characterize the epicardial region.

[0060] The method may further include configuring the system for calculating a value of the derived epicardial radiomic signature for the patient or subject. For example, the method may further include configuring the system for characterizing an epicardial region of the patient or subject by calculating a value of the derived epicardial radiomic signature for the patient or subject. The system may be configured to calculate the value of the derived epicardial radiomic signature using or based on medical imaging data of at least the epicardial region of the patient or subject. The system may be configured to calculate the value of the derived epicardial radiomic signature using or based on at least the values ​​of a second plurality of epicardial radiomic features of the epicardial region of the patient or subject.

[0061] Thus, methods may be used to derive an epicardial radiomic signature and configure a system for characterizing an epicardial region of a patient using the derived epicardial radiomic signature.

[0062] The method may further include loading computer-readable instructions onto a computer-readable memory. The instructions, when executed by the computer, cause the computer to calculate a value for the derived epicardial radiomics signature for the patient or subject. For example, the instructions may cause the computer to characterize the epicardial region of the patient or subject by calculating the value of the derived epicardial radiomics signature for the patient or subject. The instructions may cause the computer to calculate the value of the derived epicardial radiomics signature using or based on medical imaging data of at least the epicardial region of the patient or subject. For example, the instructions may cause the computer to calculate the value of the derived epicardial radiomics signature using or based on at least the values ​​of a second plurality of epicardial radiomics features of the epicardial region of the patient or subject. The instructions may be a computer program.

[0063] The system may be configured to receive as input the medical imaging data or the values ​​of the second plurality of radiomic features. The system may be configured to output (e.g., display) a calculated value of the radiomic label or output a value based on the calculated value of the radiomic label. The system may be configured to output an indication of the myocardial health of the patient. The system may be configured to output an indication of whether the patient suffers from myocardial disease. The system may be configured to output an indication of the patient's risk of suffering from a cardiac condition. The system may be configured to output an indication of the patient's risk of experiencing a stroke. The system may be a computer system.

[0064] The method may include providing instructions for configuring a system for calculating a value of a derived radiomic signature for a patient or subject.

[0065] The medical imaging data may be radiographic data. The medical imaging data may be computed tomography data.

[0066] The epicardial region may include epicardial adipose tissue.

[0067] In addition to the at least two epicardial radiomic features mentioned above, the epicardial radiomic signature of the present invention can also be calculated based on other radiomic features of the epicardial region. For example, in addition to at least two epicardial radiomic features, the epicardial radiomic signature can include other epicardial radiomic features. Therefore, it can be said that the epicardial radiomic signature is calculated based on multiple epicardial radiomic features, and the multiple epicardial radiomic features can include at least two epicardial radiomic features.

[0068] The method of the present invention may further comprise the step of calculating epicardial radiomics features from the medical imaging data.

[0069] According to a third aspect of the present invention, a system configured to perform any of the methods described above is provided. The system may be a computer system. The system may include a processor configured to perform the steps of the method. The system may include a memory loaded with executable instructions for performing the steps of the method.

[0070] According to a fourth aspect of the invention, there is provided a use of an epicardial radiomic signature for any of the above purposes, for example, characterizing an epicardial region, assessing cardiac health, detecting myocardial disease, predicting risk of developing a cardiac condition or predicting or classifying risk of experiencing a stroke. The epicardial radiomic signature may be calculated based on measured values ​​of a plurality of epicardial radiomic features of an epicardial region.

[0071] The epicardial region may be or may include epicardial tissue, for example, epicardial adipose tissue, in particular peri-atrial epicardial adipose tissue. The epicardial region may also include water and / or other soft tissue structures within the epicardial region. For example, the epicardial region may include connective tissue. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The present invention will now be described with reference to the accompanying drawings, in which:

[0073] Figure 1 The methodology used to derive and validate the epicardial radiomics signature is presented using a flow chart. Afib / AF: atrial fibrillation; CV: cross-validation; CABG: coronary artery bypass grafting; POAF: postoperative atrial fibrillation.

[0074] Figure 2 demonstrates various aspects of the principal component analysis of the radiomic features of peri-atrial adipose tissue. Figure 2(a) shows a scree plot of the percentage of variation explained by the 86 first principal components, accounting for 99.5% of the variation in study group A (843 radiomic features from 310 patients). Figure 2(b) shows a component plot of the three main principal components. Figure 2(c) shows the principal components of the radiomics of peri-atrial adipose tissue identified as independent predictors of atrial fibrillation in a logistic regression with backward elimination. Figure 2(d) shows a correlation plot of the inter-correlations between the 843 quantitative radiomic features of peri-atrial adipose tissue in the same patients. Figure 2(e) shows a Manhattan plot of the categorical values ​​of the radiomic features of peri-atrial adipose tissue for atrial fibrillation in a nested case-control cohort of study group A.

[0075] Figure 3 shows unsupervised hierarchical clustering of selected radiomic features and radiomic correlation heatmaps. Figure 3(a) shows unsupervised hierarchical clustering of patients (individuals) in group A using 33 filtered radiomic features significantly associated with atrial fibrillation. Different radiomic features are represented on the X-axis and individual patients (observations) are represented on the Y-axis. The row dendrogram identifies two different clusters of patients. Figure 3(b) shows a correlation heatmap that plots the strength of the mutual correlations between the 33 selected radiomic features and shows the clustering of the selected radiomic features by using the squared Euclidean distance between the selected radiomic features. P values ​​were calculated based on chi-square.

[0076] Figure 4 illustrates various aspects of the machine learning approach for identifying the radiomic fingerprint of atrial fibrillation on peri-atrial fat tissue and hence the myocardial phenotype. Figure 4(a) illustrates the recursive feature elimination used to select the optimal number and best features to be included in the machine learning algorithm. Figure 4(b) illustrates the model training by 5-fold cross validation repeated 3 times. Figure 4(c) illustrates the performance of various machine learning algorithms for developing radiomic signatures for atrial fibrillation classification in the training dataset. Figures 4(d) and (e) illustrate the application of the developed signature (by using 15 epicardial radiomic signatures of peri-atrial fat) for atrial fibrillation classification in the (d) training dataset and (e) validation (test) dataset and the associated confusion matrices. ROC: receiver operating characteristic; Sens: sensitivity; Spec: specificity.

[0077] Figure 5 demonstrates validation of a developed epicardial radiomics signature (score) for atrial biology and risk of atrial fibrillation. Figure 5(a) demonstrates how the developed epicardial radiomics signature from Group A was applied to the Group B cohort to characterize peri-atrial adipose tissue and explore associations with atrial gene expression profiles and atrial redox state. Scores less than 6 (calculated values ​​of the signature) were associated with a low risk of post-operative atrial fibrillation, and scores of 6 or higher were associated with a high risk of post-operative atrial fibrillation. Figure 5(b) shows that high peri-atrial epicardial radiomics scores in Group B patients were associated with gene expression profiles of atrial myocardium. Figure 5(c) shows that a peri-atrial epicardial radiomic score of 6 or higher (right, darker shading) is associated with significantly increased atrial expression of collagen (COL1A1) compared to a score below 6 (left, lighter shading), and Figure 5(d) shows that a peri-atrial epicardial radiomic score of 6 or higher (right, darker shading) is associated with significantly increased superoxide (O2-) production compared to a score below 6 (left, lighter shading). Figure 5(e) shows that in the same patient, there was a significant association between the peri-atrial adipose tissue epicardial radiomic score (right, darker shading = 6 or higher; left, lighter shading = below 6) and the incidence of postoperative atrial fibrillation (POAF). Figure 5(f) shows a density plot of the correlation between the peri-atrial epicardial radiomic score and the probability of postoperative atrial fibrillation (POAF). FIG5( g ) shows the Kaplan-Meier curves of the incidence of POAF in the peri-atrial epicardial radiomics scoring subgroups.

[0078] Figure 6 demonstrates the development of the Atriomic stroke algorithm. Figure 6(a) shows a coronary computed tomography angiography image captured at a level just inferior to the aortic valve. This slice shows a manual segmentation of the left atrium (LA) and three segmentations of the left atrial (LA) peri-tissue (masks of all voxels with attenuation from -190HU to +150HU) and anatomical annotations (letters AD). The three peri-LA tissue reservoirs (A, B, C) are segmented volumes from which radiomic features are subsequently extracted. The LA segment (D) is used for deep learning automation of the segmentation process. A: atrial septum ROI; B: LA anterior wall ROI; C: left atrial appendage ROI; D: LA ROI; E (circled): pericardium. ROI: region of interest. Figure 6(b) shows screen captures of two patients showing the cardiac anatomy as seen on CTA scans (top row) and then fed into the deep learning model (second row) the ground truth, which is a manual segmentation of the LA, left atrial appendage region, LA anterior wall region, and atrial septum region, each with different shading. The predicted images (bottom row) represent the learned segmentation as the output of the deep learning model. To compute this model, a 3D convolutional neural network with 11 layers deep and dual paths to combine multiple scaling was used to perform voxel-wise segmentation of the regions of interest in the scans. CTA: computed tomography angiography. Figure 6(c) shows a plot showing the excellent agreement between the manual left atrium segmentation (ground truth) and the automatically segmented left atrium (deep learning). The number of segmentations performed was 286. CCC: consistency correlation coefficient. Figure 6(d) shows the radiomic fingerprint (signature) used to identify stroke on tissue surrounding the LA using machine learning - recursive feature elimination to select the optimal number and best features to include in the machine learning algorithm, which revealed that out of 843 radiomic features, a set of 16 features maximized the accuracy of the Atriomic algorithm to predict stroke. On the right, the machine learning schematic shows the model training performed through 5-fold cross validation (CV) repeated 3 times. Figure 6(e) shows the incremental prognostic value of the Atriomic algorithm beyond CHA2DS2-VASc stroke risk prediction. Presented are the ROC curves and AUCs of two nested models for distinguishing stroke in the CRISP-CT cohort. Line A represents the Atriomic algorithm as a radiomic signature to predict stroke in the CRISP-CT cohort. Line B represents CHA for stroke risk prediction. 2 DS 2-VASc score. The Atriomic algorithm curve is the best performing model among the 12 experimental models and combines radiomic labels from two merged peri-left atrial segmentations. The last segment utilized in this model contains adipose tissue and connective tissue of the atrial septum and the anterior wall of the left atrium. The attenuation mask applied was -190 to +150. AUC = Area Under the Curve. Figure 6(f) shows the relationship between the Atriomic stroke algorithm and atrial biology in 86 patients undergoing cardiac surgery recruited in the OX-HVF cohort. The high-risk Atriomic stroke spectrum is associated with pathological gene expression profiles in atrial myocardial tissue samples, shown here as fibrosis-related collagen (COL1A1) and the pro-inflammatory gene tumor necrosis factor alpha (TNFa). The box plot shows that individuals identified as high-risk for stroke based on the Atriomic stroke algorithm in CRISP-CT have higher expression of inflammatory and pro-fibrotic genes in the atrial appendage tissue collected during surgery. DETAILED DESCRIPTION

[0079] The inventors have discovered that an epicardial radiomic signature (otherwise referred to as a "score" or "index") calculated based on two or more (i.e., multiple) different radiomic signatures of ECR ​​(i.e., epicardial radiomic signatures) can be used to indirectly evaluate or assess cardiac (particularly myocardial) health, as well as to detect myocardial disease or predict or classify the risk of stroke (particularly ischemic stroke). For example, the radiomic signature of the present invention indicates or is associated with cardiac (e.g., myocardial) health or disease, and can therefore be used to indirectly and non-invasively assess, evaluate or characterize the cardiac health or disease, e.g., myocardial fibrosis, myocardial redox state (particularly oxidative stress), myocardial inflammation, and myocardial gene expression patterns. The radiomic signature of the present invention can indicate or be associated with the risk of future stroke occurrence, and can therefore be used to indirectly and non-invasively assess the risk of future stroke occurrence. Generally, as used herein, "association" can mean "statistically associated", such as "statistically significantly associated". Therefore, the label of the present invention can be used to predict the risk of suffering from a cardiac condition or determine whether a patient suffers from the cardiac condition, such as arrhythmia (e.g., atrial fibrillation), ischemic heart disease, heart failure and / or cardiomyopathy, which is generally associated with myocardial disease, specifically myocardial fibrosis. The label of the present invention can be used to predict the risk of a patient experiencing or suffering from a stroke, or to identify a patient at risk of a stroke, for example, a patient at high risk of suffering from a stroke. Specifically, the label can be used to determine whether a patient suffers from myocardial disease or a cardiac condition associated with myocardial disease. Therefore, the label can be used in the risk stratification of heart disease. Therefore, the radiomics label of the present invention is preferably calculated based on two or more radiomics features of ECR, and provides a tool for non-invasively characterizing or phenotyping ECR, and is therefore also used to indirectly phenotype or evaluate the health of cardiac tissue (such as myocardium, for example, atrial myocardium).

[0080] If an individual, e.g., a patient, is identified as having a high risk of having a stroke based on the calculated values ​​of the radiomic signature, a treatment (e.g., a drug) can be administered to the patient to reduce their risk of a stroke. For example, the treatment can include prescribing a drug to reduce the risk of a stroke, such as an anticoagulant or antithrombotic drug.

[0081] The present invention takes advantage of the fact that cardiac tissue and adjacent ECR, specifically tissue within the ECR, such as adipose tissue, interact in a bidirectional manner. For example, cardiac dysfunction can lead to fatty infiltration of the myocardium through adipogenesis induced by natriuretic peptide signaling. Vice versa, myocardial disease or dysfunction can induce changes in the surrounding epicardial adipose tissue. Specifically, the present invention takes advantage of the effect of this interaction on the texture of the ECR tissue (e.g., spatial heterogeneity or variability), and thus the radiomics signature of the present invention can be constructed to provide a measure of the texture of the ECR or epicardial tissue.

[0082] The epicardial radiomic signature of the invention can be used alone to characterize ECR or provide diagnostic or prognostic information, or can be combined with existing models, such as models that include demographic characteristics and conventional risk factors.

[0083] The epicardial region (ECR) refers to an area or volume adjacent to the heart (and outside the heart), such as adjacent to the myocardium. The ECR may be an area or volume of epicardial tissue (ECT) or may include or consist of ECT. The epicardial tissue is a tissue located adjacent to the heart (i.e., myocardium) and located within the pericardium (i.e., surrounded by the pericardium). Tissues are complex biological structures and may include cells (e.g., adipocytes, neurons, etc.) as well as extracellular structures and materials (such as water) that may occupy intercellular spaces. Specifically, the ECR may be an area of ​​epicardial adipose tissue (EAT) or may include or consist of EAT. Therefore, the ECR may alternatively be referred to as an area or volume of EAT. The epicardial region may be a peri-atrial region (i.e., adjacent to the atria of the heart). Specifically, the epicardial region may be an area of ​​peri-atrial adipose tissue or may include or consist of peri-atrial adipose tissue, such as peri-atrial epicardial adipose tissue. In the case of an embodiment involving stroke, the ECR preferably includes or consists of one or more peri-atrial regions, preferably one or more peri-left atrial regions. Specifically, the ECR preferably includes or consists of an epicardial region adjacent to the atrial septum (i.e., the atrial septal epicardial region) and / or an area immediately in front of the left or right atrium (i.e., the left atrial anterior epicardial region). In the case where the epicardial region is mentioned herein with reference to an anatomical region of the heart itself (e.g., the atrial septum), it should be understood that this refers to an epicardial region adjacent to the anatomical region of the heart, for example, an epicardial region in which epicardial fat is found. Moreover, in the case of referring to the left atrium, it should be understood that this can also refer to the right atrium. This is because it has been found that the radiomic features characterizing these regions can maximize the association of the label with stroke risk. In addition, the ECR can also include connective tissue in addition to adipose tissue, particularly in embodiments involving stroke. In the case used herein, the term epicardial region can be used interchangeably with the region of interest, wherein the region of interest includes or consists of the epicardial region.

[0084] The present invention adopts radiomics methods. Radiomics is an imaging field in which data characterization algorithms are used to extract a large amount of quantitative information from imaging data. The resulting features (referred to as radiomic features) range from simple volume, shape-related or first-order statistics (such as mean or median attenuation) to second-order and higher-order statistics describing the texture of the segmented volume or region and the spatial relationship of voxels with similar or different attenuation values. Such features can identify imaging patterns with significant clinical value that cannot be identified by the naked eye, and have the potential to maximize the diagnostic yield of non-invasive ECR phenotyping.

[0085] The signature of the present invention is derived and calculated based on radiomic features, such as radiomic features extracted from medical imaging data. Specifically, the medical imaging data from which the radiomic features are extracted may correspond to at least an epicardial region (ECR), for example, peri-atrial epicardial adipose tissue, and optionally also to the heart and / or other tissues adjacent to or surrounding the ECR. As used herein, the term "epicardial radiomic signature" or "ECR radiomic signature" refers to a radiomic signature calculated based on at least two epicardial radiomic features, wherein the epicardial radiomic feature is a radiomic feature calculated based on medical imaging data of at least one epicardial region. Medical imaging data typically includes radiodensity (or attenuation) values ​​for multiple voxels of interest (in this case, the ECR and optionally also adjacent tissues), typically expressed in Hounsfield units (HU).

[0086] The medical imaging data is preferably computed tomography (CT) data, but other forms of medical imaging data (e.g., radiographic data) that provide attenuation (or radiodensity) data for voxels of an imaging region, such as three-dimensional computed laminography data, may be used instead. Typically, the medical imaging data used in the present invention is three-dimensional imaging data. Throughout the following, where reference is made to CT or another medical imaging technique, it should be understood that other suitable medical imaging techniques may be used alternatively.

[0087] ECR can include only voxels whose radiodensity (or attenuation) falls within a given or predetermined range and / or is positioned within a defined region. For example, ECR can be identified by an operator, for example, by manual contour or definition. The operator can identify ECR by checking imaging data (e.g., CT images). Therefore, ECR can include voxels that are only positioned within a defined region. Alternatively, ECR can include voxels that are only positioned within a given or predetermined distance from the outer surface of the heart (e.g., the outer surface of the heart muscle (myocardium)). The given distance can be a set value or a fixed value, such as about 5mm. ECR can be positioned between the left and right pulmonary veins. Alternatively or in addition to identifying the spatial range of ECR, ECR can be identified by applying a radiodensity (or attenuation) mask to the data and identifying PCT as including only those voxels whose radiodensity is within a given or predetermined range. For example, ECR can include only those voxels whose radiodensity is in the Hounsfield unit range of about -190HU to about +30HU. Specifically, the ECR can be defined as including only voxels whose radiodensity is in the Hounsfield unit range of about -190Hu to about -30HU. This range of attenuation values ​​generally corresponds to the radiodensity of adipose tissue. However, other ranges can be used or included, for example, about -30 to about +30 Hounsfield units, which generally correspond to the radiodensity of water. Specifically, in the case where the label predicts a stroke, the range can be wider and can cover other epicardial tissues besides fat, such as epicardial connective tissue. Therefore, the range can include voxels with a radiodensity higher than -30HU. In other words, the ECR can include voxels with a radiodensity higher than (and lower than) -30HU. For example, the range can be about -190HU to about +150HU. For example, the range can at least cover the range of about -190HU to about +150HU.

[0088] Preferably, an automatic algorithm is used to identify the ECR. Machine learning, specifically deep learning, can be used to train the algorithm to identify the ECR. For example, the algorithm may have been trained using medical imaging data in which the ECR has been manually identified and segmented. The advantage of using an automatic segmentation algorithm is that the entire process can be automatic and performed by a computer. An automatic ECR region identification or segmentation algorithm can identify sub-regions of the ECR, and the sub-regions can be combined to identify the ECR of interest. For example, the ECR can identify the atrial septum (epicardium) region and the left (or right) pre-atrial (epicardium) region, and these two regions can be merged to identify the ECR or region of interest.

[0089] The ECR can be segmented before calculating the radiomic features, and the radiomic features can be calculated from the segmented data. The segmented volume or region corresponds to the ECR, and the segmentation can remove the data corresponding to the voxels outside the ECR. Therefore, the segmentation can be achieved by identifying the ECR as described above, and then removing any voxels from the data identified as not being part of the ECR, for example, those corresponding to surrounding or adjacent tissue voxels. The segmented ECR can be extracted and used to calculate the radiomic features. For example, as described above, the segmentation can be performed by an automatic algorithm.

[0090] Computer programs or software can be used to perform calculations based on radiomic features of medical imaging data. There are various commercially available software packages for this purpose, such as 3D Slicer (available from http: / / www.slicer.org; See Fedorov, A. et al., 3D Slicer as animage computing platform for the Quantitative Imaging Network (Magn Reson Imaging) 30, 1323-1341, doi: 10.1016 / j.mri.2012.05.001 (2012)). Radiomic features can be shape-related statistics, first-order statistics or texture statistics (e.g., second-order and high-order statistics). Shape-related radiomic features and first-order radiomic features can be calculated using the original radiodensity (HU) value of the ECR voxel. To compute texture features (e.g., gray-level co-occurrence matrix [GLCM], gray-level dependency matrix [GLDM], gray-level run-length matrix [GLRLM], gray-level size zone matrix [GLSZM], and neighboring gray-level tone difference matrix [NGTDM], Tables R1-R7), the ECR voxel radiodensity or attenuation values ​​are preferably discretized into multiple bins, preferably 16 bins, preferably of equal width (e.g., a width of ten HU) to reduce noise while allowing sufficient resolution to detect biologically important spatial variations in ECR attenuation. Discretization into 16 bins is recommended as the best way to increase the signal-to-noise ratio of images used for radiomics analysis. However, discretization into more or less than 16 bins is also possible. To implement symmetric rotationally invariant results, some or all of the radiomic features, especially texture statistics (GLCM, etc.), can be calculated in all (orthogonal) directions and then averaged (e.g., using the mean or other average of the individually calculated values ​​of the features in each of the directions).

[0091] Some or all of the radiomic features, particularly those related to first-order statistics and texture-based statistics, can also be calculated for the three-dimensional wavelet transform of the raw image data, thereby generating multiple additional radiomic feature sets, for example, as described in Guo et al. (Guo X, Liu X, Wang H et al., Enhanced CT images by the wavelet transform improving diagnostic accuracy of chest nodules) Journal of Digital Imaging 2011; 24(1): 44-9). The wavelet transform decomposes the data into high-frequency and low-frequency components. At high frequencies (shorter time intervals), the resulting wavelets can capture discontinuities, ruptures, and singularities in the raw data. At low frequencies (longer time intervals), the wavelets characterize the rough structure of the data to identify long-term trends. Therefore, wavelet analysis allows the extraction of hidden and important temporal features of the raw data while improving the signal-to-noise ratio of the imaging study. The data can be decomposed into multiple (e.g., eight) wavelet decompositions by discrete wavelet transform by passing the data through a multi-level (e.g., three-level) filter bank. At each level, the data is decomposed into high-frequency components and low-frequency components by a high-pass filter and a low-pass filter, respectively. Therefore, if a three-level filter bank is used, eight wavelet decompositions corresponding to HHH, HHL, HLH, HLL, LHH, LHL, LLH, and LLL will be generated, where H represents "high pass" and L represents "low pass". Of course, more or less than eight levels can be used alternatively to decompose the data. Such decomposition can be performed using widely available software (such as the Slicer Radiomics software package incorporating the Pyradiomics library). Optionally, radiomic features can be calculated based on original (raw) data, that is, wavelet transform is not applied. Therefore, in the case of a list, group, or cluster of radiomic features disclosed herein, it should be understood that these can be reduced to exclude those radiomic features calculated based on wavelet transform. Where a radiomic feature is calculated based on a wavelet decomposition or transformation of the data, this is indicated by a suffix that the radiomic feature has been calculated based on a wavelet decomposition (e.g., HHH for highpass, highpass, highpass). Thus, for example, "skewness LLL" indicates the radiomic feature "skewness" as calculated based on an LLL wavelet decomposition. Where no suffix is ​​present, the radiomic feature is calculated based on the original (or raw) data.

[0092] Deriving radiomic signatures

[0093] The present invention provides a method for deriving a radiomic signature for characterizing an ECR (e.g., an area of ​​epicardial adipose tissue surrounding the atrium) for example, for predicting risk or identifying risk of stroke, for predicting risk of suffering from a cardiac condition (e.g., arrhythmia), or for diagnosing or detecting a myocardial disease. The radiomic signature indicates the underlying myocardial health of the myocardium adjacent to the ECR, and specifically myocardial disease, such as fibrosis or oxidative stress, which is known to cause a variety of cardiac conditions, such as arrhythmias. Therefore, cardiac conditions known to be associated with myocardial disease, such as the presence of fibrosis and / or oxidative stress, can be used as surrogate markers of myocardial health or disease and used to derive the radiomic signature of the present invention. Therefore, a radiomic signature can be derived for multiple individuals or patients (cohorts of individuals) using medical imaging data, the multiple individuals including a first group of individuals with a cardiac (i.e., heart) condition known to be associated with myocardial disease (e.g., atrial fibrillation) or a history of the cardiac condition, and a second group of individuals without a cardiac condition, such as sinus rhythm, and preferably without a history of a cardiac condition. Similarly, a known history of stroke can be used as a substitute for the presence of cardiac symptoms, particularly where the goal is to derive a signature associated with or predicting stroke risk. Alternatively, the presence of a myocardial disease such as fibrosis can be used directly to construct the signature of the present invention. Thus, patients can be alternatively divided into two groups with or without myocardial disease. However, it is generally not possible to detect myocardial disease non-invasively, while cardiac symptoms such as arrhythmias can be detected or diagnosed non-invasively (e.g., ECG). Therefore, it is preferred to construct a signature using cardiac symptoms associated with myocardial disease rather than the myocardial disease itself.

[0094] As used herein, the term cardiac (or heart) condition is used to refer to abnormal function of the heart, e.g., arrhythmia or heart failure. Thus, cardiac conditions can often be observed using noninvasive methods, as they can be observed through dysfunction of the heart. On the other hand, myocardial disease or health refers to the underlying biology or phenotype of the myocardium itself, e.g., the composition or structure of the myocardium, and often requires invasive methods (e.g., biopsy sections) to determine if it is present.

[0095] Fibrosis can be defined as the excess deposition of extracellular matrix in the myocardium (heart muscle) due to the activation and / or proliferation of cardiac fibroblasts. Myocardial redox state refers to the balance of pro-oxidant and antioxidant molecules in the tissue and depends on the generation of reactive oxygen species and their elimination from the antioxidant defense system. Myocardial inflammation is inflammation of the myocardium and can be defined as the expression of pro-inflammatory mediators.

[0096] Methods generally involve case-control studies of (human) patients with and without cardiac symptoms or myocardial disease. Individuals with cardiac symptoms or myocardial disease (or with a history) are cases (first group), and individuals without cardiac symptoms or myocardial disease are controls (second group). Alternatively, the first group (cases) can be composed of individuals with a history of stroke, and the second group (controls) can be composed of individuals without a history of stroke. Therefore, wherever cardiac symptoms or myocardial disease are mentioned below, it should be understood that this can be a history of stroke instead. Similarly, in the case of mentioning the prediction of suffering from cardiac symptoms, myocardial disease (or similar diseases), it should be understood that this can replace the prediction of suffering from stroke or experiencing stroke, especially after recording imaging data. Therefore, the replacement label indicates or predicts suffering from cardiac symptoms, which can replace the indication or prediction of suffering from stroke. For example, an automatic algorithm is used to preferably perform case-control matching, for example, 1:1 matching, to match cases with controls. Case-control matching can be performed so that each case in the first group is matched with the corresponding control in the second group. Cases and controls can be matched on clinical demographics (e.g., age, sex, obesity status, cardiovascular risk factors), cohort, and / or technical parameters related to imaging data acquisition (e.g., tube voltage and CT scanner used).

[0097] A radiomic signature can then be developed according to a stepwise approach. First, for example, as described above, multiple radiomic features are calculated from medical imaging data for each of a plurality of individuals. The radiomic features may include a selection or all of the radiomic features as defined in Tables R1-R7, and each of the radiomic features may be calculated based on one or more wavelet transforms (or wavelet decompositions) of the original image data and / or the image data, as described above. Preferably, each of the radiomic features is calculated for the original image data and for the above-mentioned eight three-dimensional wavelet decompositions of the image data. Thus, a radiomic data set comprising measured or calculated values ​​of multiple radiomic features for each of the individuals is obtained.

[0098] In the case of developing a signature associated with or predictive of stroke, the method may include identifying and optionally segmenting a plurality of epicardial regions. In other words, the ECR may include or consist of a plurality of sub-regions, and the method may include segmenting the ECR itself and one or more sub-regions of the ECR, and extracting radiomic features for the ECR and the one or more sub-regions. Thus, the plurality of radiomic features may include radiomic features of the ECR and one or more sub-regions of the ECR.

[0099] For example, the method may include segmenting an ECR including one or more (e.g., all) of the following subregions: 1) atrial septum (epicardium) region; 2) left (or right) atrial anterior (epicardium) region (region adjacent to the anterior surface of the left atrium); and 3) left (or right) atrial appendage (epicardium) region. The subregions may also include regions resulting from merging original subregions. For example, the subregions may also include merged subregions, which include or consist of adjacent subregions, particularly subregions 1+2 and / or 2+3. The method may then include calculating the ECR and features of each ECR subregion in the ECR subregions. Preferably, the ECR includes or consists of subregions 1+2+3, and the subregions are subregions 1, 2, 3, 1+2, and 2+3.

[0100] Radiomic signatures can also be calculated for both a narrower HU range corresponding to adipose tissue (e.g., about -190HU to about -30HU) and a wider HU range corresponding to epicardial tissue other than adipose (e.g., about -190HU to about +150HU). This is useful because it increases the physiological information encapsulated in the radiomic signature and thus improves the pool of information that can be used to construct a signature. For example, different tissue types can provide different markers of cardiac health. This is particularly the case for the prediction of stroke.

[0101] Radiomic features that are not significantly associated (e.g., correlated) with cardiac symptoms or myocardial diseases above a significance threshold based on data analysis can then be removed from the plurality of radiomic features. The association of each radiomic feature with a cardiac symptom or myocardial disease can be calculated based on a receiver operating characteristic (ROC) curve analysis, specifically calculating the area under the curve (AUC) based on data from a plurality of individuals. The significance threshold is preferably about α=0.05 or less, for example, α can range from 0.001 to 0.05. The significance threshold is preferably about α=0.05. However, the significance threshold can be about α=0.04. Alternatively, the significance threshold can be about α=0.03. Alternatively, the significance threshold can be about α=0.02. Alternatively, the significance threshold can be about α=0.01. Alternatively, the significance threshold can be about α=0.005. Alternatively, the significance threshold can be about α=0.002. The end result should be to remove from the plurality of radiomic features any radiomic features that are not significantly associated with the presence of a cardiac symptom or cardiomyopathy (as determined or calculated from the data, e.g., based on data analysis). In other words, the method includes selecting those features that are significantly associated or correlated with a cardiac symptom or cardiomyopathy. This has the advantage of greatly reducing the number of radiomic features to be further processed to construct a signature, thereby reducing the complexity of subsequent signature construction and alleviating the computational burden of doing so. The reason for selecting the important radiomic features is that, because they have been shown to be independently associated with cardiac health, most likely, these important radiomic features are most likely to be combined together to provide a radiomic signature associated with cardiac health.

[0102] When determining whether a feature is statistically significantly associated with a cardiac condition or myocardial disease, statistical adjustments can be applied to correct for multiple comparisons and reduce the false discovery rate (FDR). For example, a Benjamini-Hochberg correction can be applied, for example, using a false discovery rate of about 0.10. Alternatively, a Bonferroni correction can be applied to a significance threshold. The Bonferroni correction can be applied based on the number of principal components that account for a given amount of variability in the study sample based on principal component analysis. For example, the given amount can be about 99.5%. In other words, the m value used to correct the α value (by dividing α by m, i.e., α / m) is the number of principal components that account for a given amount of variability. Therefore, principal component analysis of radiomics features can be performed on data from multiple individuals.

[0103] The remaining or "important" radiomic features (i.e., radiomic features found to be statistically significantly associated with cardiac symptoms or myocardial disease) can then be grouped or "clustered" into multiple clusters of similar or related features. The degree of correlation between features is a measure of the degree to which two radiomic features tend to vary from one another between different individuals. Pairwise correlations can be calculated using the Spearman rho coefficient or other measures of correlation (such as the Pearson correlation coefficient). Clustering can be performed, for example, using a hierarchical clustering method (such as a hierarchical clustering algorithm) to classify important radiomic features into multiple clusters. Hierarchical clustering can be performed in an unsupervised manner, that is, independent of the strength of the correlation of the radiomic features with cardiac symptoms or myocardial disease. In other words, clustering can be performed based on the strength of the correlation between features, so that radiomic features and those radiomic features that are most correlated with the radiomic features are clustered together. In particular, intra-cluster correlations can be stronger than inter-cluster correlations, that is, correlations between features within a cluster are stronger than correlations between features in different clusters. The final identification of clusters can be performed by a person (e.g., an operator) by inspecting the correlation data. For example, an operator can inspect a dendrogram and / or a two-dimensional correlation plot (or heat map) representing the hierarchical clustering of radiomic features, which plots the correlation between each radiomic feature and each other radiomic feature (and itself) in the radiomic features, and can identify clusters based on this inspection of the radiomic feature inter-correlation data. In the correlation heat map, the radiomic features can be arranged along the x-axis and y-axis of the correlation plot in an order determined according to the hierarchical clustering (i.e., where the features are positioned near the features that are most closely associated or related to the features). This visual inspection can be used in conjunction with the hierarchical clustering to identify appropriate clusters of radiomic features. Alternatively, clusters can be identified by hierarchical clustering alone or by visual inspection.

[0104] A radiomic signature can be constructed based on at least two (or all) of the multiple important features (i.e., features found to be significantly associated with cardiac symptoms or myocardial disease), at least two of the radiomic features being different from each other. Preferably, a radiomic signature is constructed based on at least two of the multiple important features, wherein at least two of the radiomic features are selected from different clusters. For example, a radiomic signature can be constructed based on at least one radiomic feature selected from each cluster. Other radiomic features may also be included in the initial signature to be optimized, for example, two or more radiomic features from any or all of the clusters may be included in the initial signature. However, in order to provide a signature more closely associated with cardiac symptoms or myocardial disease and thus improve the usefulness of diagnosis and prognosis, it is preferred to include at least two radiomic features, each of which is from a different cluster. This is because features from different clusters provide supplementary phenotypic information associated with ECR. Specifically, radiomic features from different clusters may be sensitive to different phenotypic features of ECR ​​because they are found not to vary in a similar manner, indicating that they are sensitive to different phenotypic features of ECR. Therefore, the radiomic signature is preferably calculated based on (the measured values ​​of) at least two radiomic features, each radiomic feature being selected from a different cluster. For example, the radiomic signature may include at least three radiomic features, each radiomic feature being selected from a different cluster. Alternatively, the initial radiomic signature may include at least four radiomic features, each radiomic feature being selected from a different cluster. Alternatively, the initial radiomic signature may include at least five radiomic features, each radiomic feature being selected from a different cluster. Alternatively, the initial radiomic signature may include at least six radiomic features, each radiomic feature being selected from a different cluster. Preferably, the initial radiomic signature may include one radiomic feature from each cluster.

[0105] Construction of a radiomic signature may involve refining or optimizing the radiomic signature, specifically using data from a subset of a cohort known as a "training" cohort. This involves refining or optimizing the signature based on the data to improve the correlation or association between the signature and the cardiac condition.

[0106] The label may include (i.e., the value of the label may be calculated using a decision tree) a decision tree, wherein the input variable is or includes the radiomic feature of the label. The target variable or outcome may be calculated by the decision tree based on the input variable. Specifically, because the radiomic feature is a continuous variable, the decision tree may be a regression tree. The decision tree algorithm uses branching options at each tree node, and repeating this process at multiple levels will produce the final branches or tree leaves. Therefore, each observation (i.e., the set of radiomic feature values ​​input into the decision tree) is assigned to the final leaf, and this provides an associated classification probability (similar to a logistic regression probability) for the presence of a disease or condition. It was found that formulating the radiomic label as a decision tree can provide the best sensitivity for assessing cardiac health. Specifically, it was found that the decision tree provided the best accuracy for distinguishing patients with cardiac symptoms or myocardial disease from patients who do not suffer from the cardiac symptoms or myocardial disease.

[0107] Alternatively, the label may include a weighted sum of the values ​​of each of the radiomic features contained in the label, and the weight of each of the radiomic features may be gradually optimized or improved. The coefficients by which each of the radiomic features is multiplied are generally referred to as beta (β) coefficients, and it is these β coefficients that may be optimized or improved.

[0108] One or more machine learning algorithms may be used to construct or derive the signature. For example, a machine learning algorithm may be used to refine the contribution of radiomic features to the signature so as to increase or maximize the association of the signature with cardiac symptoms or myocardial disease. For example, a plurality (preferably all) of the important radiomic features (i.e., features significantly associated with cardiac symptoms or myocardial disease) may be input into a first machine learning algorithm. The first machine learning algorithm may be used to identify the optimal number and identity of important radiomic features to be included in the signature, specifically to maximize their accuracy for distinguishing cardiac symptoms or myocardial disease. In other words, the construction of the radiomic signature may include a feature selection step, in which the radiomic features to be included in the signature are selected. In the feature selection step, the feature selection algorithm (e.g., a machine learning algorithm) may select a subset of radiomic features to be included in the final signature, specifically radiomic features predicted to maximize the association of the final signature with cardiac symptoms or myocardial disease. This has the advantage of reducing the complexity of optimizing the final signature because it reduces the number of radiomic features that need to be considered. The first machine learning algorithm that performs this feature selection step may use recursive feature elimination, for example, in conjunction with a random forest algorithm. However, other algorithms may be used instead. The first machine learning algorithm may be constrained to require that the resulting signature includes at least two radiomic features selected from different clusters. For example, the first machine learning algorithm may be constrained to require that the resulting signature includes at least one radiomic feature selected from each of the clusters.

[0109] Once the number and identity of radiomic features to be included in the label are identified, a second machine learning algorithm can be used to optimize the contribution of each of the features identified by the first machine learning algorithm to the label. In other words, the construction of the radiomics label can include a step of using a second machine learning algorithm to refine or optimize the radiomics label to increase or maximize its association with cardiac disease or myocardial disease. The second machine learning algorithm can be a decision tree learning algorithm (e.g., if the radiomics label includes or is a decision tree), and is preferably a gradient boosting algorithm, specifically an extreme gradient boosting algorithm. The gradient boosting algorithm is a well-known decision tree learning algorithm for classification / regression. The initial cohort (or multiple individuals) can be divided into a training cohort and a test cohort, for example, using a random seed. The training cohort can be composed of approximately 80% of the individuals from the initial cohort, and the test cohort can be composed of approximately 20% of the individuals from the initial cohort. The data of the training cohort can be used to derive or refine the label, and the data of the test cohort can be used to verify the label. For example, an internal cross-validation can be used to construct the label. The internal cross-validation can be multiple, for example, 5-fold.

[0110] Alternatively, a single machine learning algorithm can be used to construct a signature from important radiomic features, rather than in the two-step process described above. For example, a single machine learning algorithm, such as a decision tree learning algorithm with inherent feature selection, can be used to identify features to be included in the final signature and refine or optimize the signature. For example, the number of radiomic features to be included in the final signature can be pre-selected or predetermined and input as a parameter to the machine learning algorithm, or can be turned on and selected by the machine learning algorithm itself, which can also refine and optimize the radiomic signature. In other words, the machine learning algorithm can include feature selection capabilities.

[0111] Although the method described above produces a radiomic signature constructed by finding only radiomic features that are significantly associated with cardiac symptoms or myocardial disease (i.e., "important" features), the radiomic signature does not have to contain only important radiomic features. Therefore, the construction of the radiomic signature discussed above does not need to be performed using only important radiomic features, and can be performed without first selecting only important radiomic features, so that non-important features are also included in the construction of the features. Alternatively, any or all important radiomic features of the constructed signature can be replaced with radiomic features that are highly correlated or collinear with the important features (i.e., collinear equivalents). A label in which one or more important features in the important features are replaced with features that are collinear with the features will generally be performed in a similar manner to a label calculated based on only important features, because by definition, collinear features behave very similarly to each other. In fact, it is possible that replacing one or more (even all) important features in the important features with alternative features that are collinear with the replaced important features may result in a label with an improved prognostic value, and it has actually been found to be the case in some cases. This is because although the raw features are often most independently associated with the clinical endpoint, they are not necessarily the best performing features when combined into a signature.

[0112] Therefore, the method of deriving a signature may include replacing one or all of the important radiomic features with a radiomic feature that is found to be collinear with the replaced important feature. Specifically, once the important features that maximize the association of the signature with cardiac symptoms or myocardial disease are identified, any one or all of these important features can be replaced with features that are collinear with the replaced important features. Therefore, the radiomic signature may include at least two of the important radiomic features or their collinear equivalents (i.e., calculated based on the at least two radiomic features). For example, the radiomic signature may include one or more important radiomic features and one or more collinear equivalents of other important radiomic features. Preferably, at least two important radiomic features are selected from different clusters (or the replacement features correspond to important radiomic features belonging to different clusters). Preferably, the radiomic signature includes at least two of the important radiomic features (or their collinear equivalents) that are found to maximize the accuracy of the signature for predicting cardiac symptoms or myocardial disease.

[0113] Therefore, the method of deriving a label may include evaluating the pairwise correlation between radiomic features and identifying a group of radiomic features that are correlated or collinear with important radiomic features. The measured values ​​of the radiomic features of multiple individuals are used to calculate the correlation between radiomic features. Collinear radiomic features can be identified as radiomic features that are correlated to each other at least equal to a given correlation threshold in individuals. The correlation threshold is preferably applied to both positive and negative correlations, for example, the correlation threshold can be expressed as a modulus. Pairwise correlations can be calculated using the Spearman rho coefficient, and the correlation threshold can be at least about |rho|=0.75, for example, about |rho|=0.75, so that all paired radiomic features that are correlated to each other at the level of |rho|≥0.75 are considered to be collinear with each other. Alternatively, the correlation threshold can be at least about |rho|=0.9, for example, about |rho|=0.9. Alternatively, the correlation threshold can be at least about |rho|=0.9, for example, about |rho|=0.9. As will be readily understood in the art, correlation or collinearity is a measure of how closely two radiomic features vary together from one individual to the next, and can be calculated based on measured radiomic feature values ​​for multiple individuals.

[0114] A radiomic signature can then be constructed from at least two radiomic features selected from different groups of collinear features. In other words, a signature can be composed of at least two different important radiomic features or replace a radiomic feature that is collinear with two different important radiomic features.

[0115] As mentioned above, a label may contain a weighted sum of the calculated values ​​of multiple radiomic features. Labels may also contain other terms, such as addition or subtraction of constants, or multiplication of factors. However, typically, if a label contains a weighted sum, it will be linearly related to the weighted sum of the radiomic feature values ​​in some way.

[0116] Radiomics labels can be of the following forms or contain terms (e.g., labels can be calculated based on functions of containing terms):

[0117] A ±Σb i rf i

[0118] Where A is a constant (can be zero or non-zero), b i is the weighting coefficient (or β parameter) of radiomic feature i, and rf i is the measured value of radiomic feature i.

[0119] However, preferably, a decision tree is used to calculate the value of the radiomics signature. In this case, the decision tree can output a predicted probability (P) of the presence of a cardiac symptom or myocardial disease based on the measured values ​​of the radiomics features input into the decision tree. Therefore, the final value of the label can be related to P in some way. In other words, the radiomics signature can be simply P, or it can be calculated using the value P. For example, the label can also contain other terms, such as addition or subtraction of a constant, or multiplication of a factor or constant. However, typically, the label will be linearly related to the output of the decision tree. For example, a radiomics signature can be calculated as A×P, where A is a constant. A can be, for example, 10.

[0120] Typically, in the above methods, bivariate associations between radiomic features can be assessed by the nonparametric Spearman rho (ρ) coefficient.

[0121] Radiomics signature

[0122] The ECR radiomics signature of the present invention is calculated based on the measured values ​​of the radiomics features obtained from the medical imaging data. In particular, the ECR radiomics signature is preferably calculated based on at least two radiomics features.

[0123] To improve the prognostic and diagnostic value of the signature, as described above, the signature is preferably calculated based on at least two different radiomic features selected from different clusters of similar or related features. This reduces redundancy and improves the diversity of information included in the signature calculation, because features from different clusters relate to different texture aspects of the ECR.

[0124] Four clusters (AD) were identified using a hierarchical clustering algorithm (see Example). Table 1 (and Figure 3b ) identifies members of four clusters. The radiomic signature may include at least two of the radiomic features from Table 1. Advantageously, the radiomic signature may be calculated based on at least two of the radiomic features selected from the clusters AD identified in Table 1, the at least two radiomic features being selected from different clusters. Preferably, the radiomic signature may be calculated based on at least one radiomic feature selected from each of the clusters AD identified in Table 1.

[0125] Table 1: Clustering of radiomic features

[0126]

[0127] Fifteen radiomic features were found to maximize the association of radiomic signatures with cardiac symptoms, and these radiomic features are shown in Table 2. The radiomic signature can be advantageously calculated based on at least two radiomic features from the radiomic features in Table 2. Preferably, the radiomic signature is calculated based on at least two radiomic features, each of the at least two radiomic features being selected from a different cluster AD. Further preferably, the radiomic signature can be calculated based on at least one radiomic feature of each cluster selected from the clusters AD identified in Table 2. In order to maximize the association of the radiomic signature with cardiac symptoms, the radiomic signature can be calculated based on all 15 radiomic features listed in Table 2.

[0128] Table 2: Further optimized radiomics feature clustering

[0129]

[0130] As previously described, the important radiomic features of Tables 1 and 2 can be replaced with other radiomic features (i.e., collinear equivalents) that are related to or collinear with the replaced important radiomic features to obtain a signature with similar diagnostic and prognostic usefulness. Therefore, a radiomic signature can be calculated based on (i.e., including) at least two radiomic features selected from the radiomic features of Table 3. Each of the groups identified in Table 3 contains one of the 15 important radiomic features that were found to maximize the association of the signature with a cardiac condition and those radiomic features that are calculated to be collinear with the important features to a degree of at least |rho|=0.75, where rho is Spearman's rho. Therefore, a radiomic signature can be constructed as described above, but one or more of the important radiomic features of Table 2 are replaced with radiomic features that are collinear with the features, as shown in Table 3. For example, a radiomic signature can be calculated based on at least two radiomic features, each of which is selected from a different group of Table 3. Specifically, the radiomic signature can be calculated based on at least two radiomic features selected from a group corresponding to important features belonging to different clusters AD. Further preferably, the radiomic signature can be calculated based on at least one radiomic feature selected from each cluster in the clusters AD identified in Table 2 or a collinear equivalent thereof. In other words, the radiomic signature can be calculated based on at least one radiomic feature selected from each cluster in the clusters AD identified in Table 3 below.

[0131] Table 3: Groups of radiomic features collinear with significant features

[0132]

[0133]

[0134]

[0135]

[0136]

[0137]

[0138]

[0139]

[0140]

[0141]

[0142]

[0143]

[0144]

[0145]

[0146]

[0147]

[0148]

[0149]

[0150]

[0151]

[0152]

[0153]

[0154]

[0155]

[0156]

[0157]

[0158]

[0159]

[0160]

[0161]

[0162]

[0163]

[0164]

[0165]

[0166]

[0167]

[0168]

[0169] The groups identified in Table 3 can be reduced to include only those radiomic features (i.e., one of the 15 important features identified by the feature selection algorithm) that correlate with the original important features of the group by at least |rho|=0.800 (this includes the important feature itself, which, by definition, has a correlation with itself of rho=1). For example, the group can be reduced to those features that correlate with the important features of the group by at least |rho|=0.850. For example, the group can be reduced to those features that correlate with the important features of the group by at least |rho|=0.900. For example, the group can be reduced to those features that correlate with the important features of the group by at least |rho|=0.950.

[0170] In addition to calculating a radiomic signature based on at least two radiomic features from different clusters or groups, the radiomic signature may also be calculated based on additional radiomic features. For example, a radiomic signature may include more than one radiomic feature from any given cluster or group, or may include radiomic features that are not included in any of the clusters or groups. Therefore, it can be said that a radiomic signature is calculated based on a plurality of radiomic features, and the plurality of radiomic features may include at least two of the radiomic features mentioned above.

[0171] Stroke

[0172] In order to maximize the correlation with stroke, the radiomic signature is calculated again based on the measured values ​​of the radiomic features obtained from the medical imaging data. Specifically, the radiomic signature is preferably calculated based on at least two radiomic features.

[0173] In order to improve the prognostic and diagnostic value of the signature, as already described, the signature is preferably calculated based on at least two different radiomic features selected from different clusters of similar or related features. This reduces redundancy and improves the diversity of information included in the signature calculation, since features from different clusters relate to different textural aspects of the epicardial tissue.

[0174] Four clusters (AD) are identified using a hierarchical clustering algorithm. Table 1b identifies the members of the four clusters. The radiomics signature may include at least two of the radiomics features from Table 1b. Advantageously, the radiomics signature may be calculated based on at least two of the radiomics features selected from the clusters AD identified in Table 1b, the at least two radiomics features being selected from different clusters. Preferably, the radiomics signature may be calculated based on at least one radiomics feature selected from each of the clusters AD identified in Table 1b.

[0175] Table 1b: Clustering of radiomic features

[0176]

[0177] 16 radiomic features were found to maximize the association of radiomic signatures with ischemic stroke, and these radiomic features are shown in Table 2b. Radiomic signatures can be advantageously calculated based on at least two radiomic features from the radiomic features in Table 2b. Preferably, the radiomic signature is calculated based on at least two radiomic features, each of the at least two radiomic features being selected from different clusters AD. Further preferably, the radiomic signature can be calculated based on at least one radiomic feature of each cluster selected from the clusters AD identified in Table 2b. In order to maximize the association of radiomic signatures with stroke, radiomic signatures can be calculated based on all 16 radiomic features listed in Table 2b.

[0178] Table 2b: Further optimized radiomics feature clustering

[0179]

[0180] As previously described, the important radiomic features of Table 1b and Table 2b can be replaced with other radiomic features (i.e., collinear equivalents) that are related to or collinear with the replaced important radiomic features to obtain labels with similar diagnostic and prognostic usefulness. Therefore, a radiomic signature can be calculated based on (i.e., including) at least two radiomic features selected from the radiomic features of Table 3b. Each of the groups identified in Table 3b contains one of the 16 important radiomic features that were found to maximize the association of the signature with stroke and those radiomic features that are calculated to be collinear with the important features for at least |rho|=0.75, where rho is Spearman's rho. Therefore, a radiomic signature can be constructed as described above, but one or more of the important radiomic features of Table 2b are replaced with radiomic features that are collinear with the features, as shown in Table 3b. For example, a radiomic signature can be calculated based on at least two radiomic features, each of which is selected from different groups of Table 3b. Specifically, the radiomic signature can be calculated based on at least two radiomic features selected from a group corresponding to important features belonging to different clusters AD. Further preferably, the radiomic signature can be calculated based on at least one radiomic feature selected from each cluster in the clusters AD identified in Table 1b or its collinear equivalent. In other words, the radiomic signature can be calculated based on at least one radiomic feature selected from each cluster in the clusters AD identified in Table 3b below.

[0181] Table 3b: Groups of radiomic features collinear with significant features

[0182]

[0183]

[0184]

[0185]

[0186]

[0187]

[0188]

[0189]

[0190]

[0191]

[0192]

[0193]

[0194]

[0195]

[0196]

[0197]

[0198]

[0199]

[0200]

[0201]

[0202]

[0203]

[0204]

[0205]

[0206]

[0207]

[0208]

[0209]

[0210]

[0211]

[0212]

[0213]

[0214] The groups identified in Table 3b can be reduced to include only those radiomic features (i.e., one of the 16 important features identified by the feature selection algorithm) that correlate with the original important features of the group by at least |rho|=0.800 (this includes the important feature itself, which, by definition, has a correlation with itself of rho=1). For example, the group can be reduced to those features that correlate with the important features of the group by at least |rho|=0.850. For example, the group can be reduced to those features that correlate with the important features of the group by at least |rho|=0.900. For example, the group can be reduced to those features that correlate with the important features of the group by at least |rho|=0.950.

[0215] In addition to calculating a radiomic signature based on at least two radiomic features from different clusters or groups, the radiomic signature may also be calculated based on additional radiomic features. For example, a radiomic signature may include more than one radiomic feature from any given cluster or group, or may include radiomic features that are not included in any of the clusters or groups. Therefore, it can be said that a radiomic signature is calculated based on multiple radiomic features, and the multiple radiomic features may include at least two radiomic features.

[0216] For radiomic features associated with or predictive of stroke, the signature preferably includes (i.e., calculated based on) two or more radiomic features of an ECR that includes or consists of a peri-atrial region, specifically a peri-left atrial region, as these regions have been found to be most closely associated with stroke. Specifically, the ECR preferably includes or consists of a mid-atrial septal epicardial region and a left (or right) atrial pre-epicardial region. The radiomic features of these epicardial subregions have been found to be most strongly predictive of stroke, and therefore provide a more reliable and useful signature for predicting stroke.

[0217] The epicardial radiomic signature can be calculated based on the measured values ​​of multiple epicardial radiomic features of the epicardial region. The epicardial radiomic features can each be calculated for the epicardial region as a whole, or can be calculated for a sub-region of the epicardial region (e.g., the atrial septum and / or the anterior left atrial region). All epicardial radiomic features can be calculated for the same region or sub-region, but this is not required, and different radiomic features can be calculated for different sub-regions.

[0218] Likewise, for stroke prediction, the radiomic signature is preferably calculated for voxels within a wider attenuation (HU) range corresponding to fat and other epicardial tissues (e.g., the attenuation range includes voxels with HU attenuation above -30HU, e.g., about -190 to about +150). In other words, the epicardial region is preferably calculated for the epicardial region comprised of or consisting of voxels within this wider HU range. This is because tissues other than fat have been found to predict stroke, and therefore, when a wider HU range corresponding to epicardial tissues other than fat is used, it improves the predictive power of the signature.

[0219] Each radiomic signature of the radiomic signature of the present invention provides a direct way to characterize peri-atrial tissue using medical imaging data. Because each radiomic signature of the radiomic signature of the present invention is based on a relatively small number of possible radiomic features that can be measured, the signature is easy to calculate and understand, and clinicians can better understand its physiological significance.

[0220] Each radiomic signature in the radiomic signature of the present invention provides a direct way to characterize ECR using medical imaging data. Because each radiomic signature in the radiomic signature of the present invention is based on a relatively small number of possible radiomic features that can be measured, the signature is easy to calculate and understand, and clinicians can better understand its physiological significance.

[0221] system

[0222] The method of the present invention can be performed on a system such as a computer system. Therefore, the present invention also provides a system configured or arranged to perform one or more methods in the method of the present invention. For example, the system may include a computer processor configured to perform one or more methods or steps of the method in the method of the present invention. The system may also include a computer readable memory loaded with executable instructions for performing the steps of any method in the method of the present invention.

[0223] Specifically, the method for deriving radiomics labels can be performed on such systems, and therefore such systems are provided according to the present invention. For example, the system can be configured to receive and optionally store a data set, the data set including the values ​​of multiple radiomics features of ECR ​​obtained from medical imaging data for each individual in a plurality of individuals. The system can be configured to construct (e.g., derive and verify) radiomics labels using such data sets according to the method of the present invention.

[0224] Alternatively, the system can be configured to perform a method of characterizing ECR or assessing cardiac health. Specifically, the present invention provides a system for characterizing ECR or assessing cardiac health using medical imaging data of a subject. The system can be configured to calculate the value of a radiomics signature of the ECR using the medical imaging data. The radiomics signature can be calculated based on the measured values ​​of at least two radiomics features of the ECR, and the measured values ​​of at least two radiomics features can be calculated based on the medical imaging data.

[0225] The system may also be configured to compute radiomic features from medical imaging data, as described in more detail above.Thus, the system may be configured to receive and optionally store medical imaging data, and process the imaging data to compute radiomic features.

[0226] Definition of radiomic features

[0227] The definitions of radiomic features referred to herein are generally well understood within the radiomics field simply by reference to their names. However, for ease of reference, the definitions of the features used herein are provided in Tables R1 to R7 below. The radiomic features in Tables R1 to R7 are defined based on the radiomic features used by the Pyradiomics package (http: / / pyradiomics.readthedocs.io / en / latest / features.html, see van Griethuysen, JJM, Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., Beets-Tan, RGH, Fillon-Robin, JC, Pieper, S., Aerts, HJWL (2017). Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Research, 77(21), e104-e107. https: / / doi.org / 10.1158 / 0008-5472.CAN-17-0339). Most of the features defined in Tables R1 to R7 conform to the feature definitions described by the Imaging Biomarker Standardization Initiative (IBSI), which can be found in Zwanenburg et al., (2016) (Zwanenburg, A., Leger, S., Vallières, M. and S. (2016). Image biomarker standardisation initiative - feature definitions In eprint arXiv:1612.07003 [cs.CV]). In the case where the definitions provided below do not exactly conform to the IBSI definitions, it will be appreciated that either definition may be used in accordance with the present invention. Ultimately, the precise mathematical definition of radiomic features is not critical, as slight modifications will not affect the general properties of the image measured by each of the features. Therefore, slight modifications to features (e.g., addition or subtraction of constants or scaling) and alternative definitions of features are intended to be encompassed by the present invention.

[0228] a. First-order statistics

[0229] These statistics describe the central tendency, variability, uniformity, asymmetry, skewness, and amplitude of attenuation values ​​in a given region of interest (ROI), while ignoring the spatial relationship of individual voxels. Thus, the statistics describe quantitative and qualitative characteristics of the entire ROI (ECR). A total of 19 features were calculated for each of the eight wavelet transforms and the original CT image, as follows:

[0230] make:

[0231] X is a set of N contained in the region of interest (ROI) p The attenuation or radiodensity value of the voxel (e.g., in HU)

[0232] P(i) is a g A first-order histogram of discrete intensity levels, where N g is the number of non-zero bins, equally spaced from 0 by a certain width.

[0233] p(i) is the normalized first-order histogram and is equal to

[0234] c is a value that varies the intensity to prevent negative values ​​in X. This ensures that the voxels with the lowest grayscale values ​​contribute the least to the energy, rather than the voxels with grayscale intensity closest to 0. Since the HU range of adipose tissue (AT) within the ECR (-190 to -30 HU) does not include zero, c can be set to c=0. Therefore, higher energies correspond to less radiodense AT, and therefore higher lipophilic content. However, other values ​​of c can be used.

[0235] ∈ is an arbitrarily small positive number (about 2.2×10 -16 )

[0236] Table R1: First-order radiomic features for ECR characterization

[0237]

[0238]

[0239]

[0240] b. Shape-related statistics

[0241] Shape-related statistics describe the size and shape of a given ROI without taking into account the attenuation values ​​of its voxels. Since they are independent of grayscale intensity, shape-related statistics are consistent across all wavelet transformed and raw CT images and are therefore calculated only once. These are defined as follows:

[0242] make:

[0243] V is the volume of ROI, in mm 3 Unit

[0244] A is the surface area of ​​ROI, in mm 2 Unit

[0245] Table R2: Shape-related radiomic features for ECR characterization

[0246]

[0247] c. Gray Level Co-occurrence Matrix (GLCM)

[0248] In simple terms, GLCM describes the number of times a voxel of a given attenuation value i is located near a voxel of j. g ×N g The GLCM describes the second-order joint probability function of the image region constrained by the mask and is defined as P(i,j|δ,θ). The (i,j)th element of this matrix represents the number of times the combination of levels i and j occurs in two pixels in the image that are separated by a distance of δ pixels along an angle θ. The distance δ from the central voxel is defined as the distance according to the infinity norm. For δ=1, this will result in 2 neighbors for each of the 13 angles in 3D (26 connectivity), and for δ=2, 98 connectivity (49 unique angles). In order to obtain rotationally invariant results, the statistics need to be calculated in all directions and then averaged to ensure a symmetric GLCM.

[0249] make:

[0250] ∈ is an arbitrarily small positive number (about 2.2×10 -16 )

[0251] P(i,j) is the co-occurrence matrix of any δ and θ

[0252] p(i,j) is the normalized co-occurrence matrix and is equal to

[0253] Ng is the number of discrete intensity levels in the image

[0254] is the marginal row probability

[0255] is the marginal column probability

[0256] μ x For p x The average gray intensity of

[0257] μ y forp y The average gray intensity of

[0258] σ x Yes x Standard Deviation

[0259] σ y Yes y Standard Deviation

[0260] Where i+j=k, and k=2, 3, ..., 2N g

[0261] where |ij|=k, and k=0, 1, ..., N g -1

[0262] Yes x Entropy

[0263] Yes y Entropy

[0264]

[0265]

[0266] For distance weighting, the GLCM matrix is ​​weighted by the weighting factor W and then summed and normalized. A weighting factor W is calculated for the distance between neighboring voxels, where d is the distance of the associated angle.

[0267] Table R3: Gray Level Co-occurrence Matrix (GLCM) Statistics for ECR Characterization

[0268]

[0269]

[0270]

[0271] d. Grayscale Size Zone Matrix (GLSZM)

[0272] Grayscale size zones (GLSZM) describe the grayscale zones in the ROI, which are defined as the number of connected voxels sharing the same grayscale intensity. A voxel is considered connected if the distance according to the infinity norm is 1 (26 connected zones in 3D, 8 connected zones in 2D). In the grayscale size zone matrix P(i,j), the (i,j)th element is equal to the number of zones of grayscale i and size j present in the image. In contrast to GLCM and GLRLM, GLSZM is rotation-independent, where only one matrix is ​​calculated for all orientations in the ROI.

[0273] make:

[0274] N g is the number of discrete intensity values ​​in the image

[0275] N s is the number of discrete region sizes in the image

[0276] N p is the number of voxels in the image

[0277] N z is the number of regions in the ROI, which is equal to and

[0278] 1≤N z ≤N p

[0279] P(i,j) is the size region matrix

[0280] P(i,j) is the normalized size region matrix, defined as

[0281] ∈ is an arbitrarily small positive number (about 2.2×10 -16 ).

[0282] Table R4: Grayscale size zone matrix (GLSZM) statistics for ECR characterization

[0283]

[0284]

[0285]

[0286] e. Grayscale run length matrix (GLRLM)

[0287] The grayscale run length matrix (GLRLM) describes the grayscale runs of consecutive pixels with the same grayscale value, where the grayscale run is defined as the length of the number of pixels. In the grayscale run length matrix P(i,j|θ), the (i,j)th element describes the number of times a run with grayscale i and length j occurs in the image (ROI) along the angle θ.

[0288] make:

[0289] N g is the number of discrete intensity values ​​in the image

[0290] N r is the number of discrete run lengths in the image

[0291] N p is the number of voxels in the image

[0292] N z (θ) is the number of runs along angle θ in the image, equal to And 1≤N z (θ)≤N p P(i,j|θ) is the run length matrix of any direction θ

[0293] p(i,j|θ) is the normalized run length matrix, defined as

[0294] ∈ is an arbitrarily small positive number (about 2.2×10 -16 ).

[0295] By default, the value of the feature is calculated separately for each angle on the GLRLM, and the mean of these values ​​is then returned. If distance weighting is enabled, the GLRLM is weighted by the distance between neighboring voxels, and then summed and normalized. The features are then calculated on the resulting matrix. The distance between neighboring voxels is calculated for each angle using the norm specified in "weightingNorm".

[0296] Table R5: Gray-level run length matrix (GLRLM) statistics for ECR characterization

[0297]

[0298]

[0299]

[0300] f. Neighboring Gray Tone Difference Matrix (NGTDM) features

[0301] The neighboring grayscale tone difference matrix quantifies the difference between a grayscale value and the average grayscale value of its neighbors within a distance δ. The sum of the absolute differences of grayscale i is stored in the matrix. Let X gl is a set of segmented voxels, and x gl (j x , j y , j z )∈X gl For the location (j x , j y , j z ), the average grayscale of the neighborhood is:

[0302]

[0303] Where (k x , k y , k z )≠(0,0,0) and x gl (j x +k x , j y +k y , j z +k z )∈X gl

[0304] Here, W is also in X gl The number of voxels in the neighborhood of .

[0305] make:

[0306] n i is X with gray level i gl The number of voxels in

[0307] Nv,p is X gl The total number of voxels in , and is equal to ∑n i (i.e., the number of voxels with a valid region; at least 1 neighbor). N v，p ≤N p , where N p is the total number of voxels in the ROI.

[0308] p i is the gray level probability and is equal to n i / N v

[0309] is the sum of absolute differences of gray level i

[0310] N g is the number of discrete gray levels

[0311] N g,p For pi The number of gray levels when ≠0

[0312] Table R6: Neighboring Gray Tone Difference Matrix (NGTDM) for ECR Characterization

[0313]

[0314] g. Gray Level Dependency Matrix (GLDM)

[0315] The gray-level dependency matrix (GLDM) quantifies gray-level dependency in an image. Gray-level dependency is defined as the number of connected voxels that are within a distance δ and that depend on a central voxel. A neighboring voxel of gray-level j is considered to be dependent on a central voxel of gray-level i if |ij|≤α. In the gray-level dependency matrix P(i,j), the (i,j)th element describes the number of times a voxel of gray-level i appears in the image that has j related voxels in its neighborhood.

[0316] N g is the number of discrete intensity values ​​in the image

[0317] N d is the number of discrete dependency sizes in the image

[0318] N z is the number of dependent regions in the image, which is equal to

[0319] P(i,j) is the dependency matrix

[0320] P(i,j) is the normalized dependency matrix, defined as

[0321] Table R7: Gray Level Dependence Matrix (GLDM) Statistics for ECR Characterization

[0322]

[0323]

[0324] Examples

[0325] Example 1

[0326] method

[0327] Study Design

[0328] Cohort A was a nested case-control study of a cohort (cohort A) of 2,246 patients undergoing diagnostic coronary CT angiography (Cleveland Clinic, US). Patients with AF (n=155) were matched 1:1 with control subjects without a known history of AF (n=155). Both groups were matched for age, sex, cardiovascular risk factors, and scan acquisition details. This panel was used to identify radiomic features of epicardial regions independently associated with AF, specifically peri-atrial fat or adipose tissue, to form a correlated radiomic score for peri-atrial fat associated with cardiac and specifically atrial health.

[0329] Group B consisted of an independent cohort of 225 patients undergoing coronary artery bypass grafting (CABG) (cohort B). Patients underwent coronary CT angiography scans and perioperative atrial tissue samples were collected for myocardial phenotyping as described below. Patients were followed up in hospital for the development of AF. This cohort was used to: a) externally validate the developed epicardial radiomics score from group A for atrial biology, and b) explore its predictive value for the development of postoperative arrhythmias, specifically AF.

[0330] Computed tomography studies

[0331] Participants in study groups A and B underwent coronary CT angiography imaging using a standardized clinical protocol. Heart rate was optimized using intravenous beta-blockers, and glyceryl trinitrate (800 μg) was also administered sublingually to achieve maximal coronary vasodilation. Iodine-based contrast agents were administered intravenously to achieve vascular opacification and diagnosis of obstructive coronary artery disease. Prospective ECG-triggered axial acquisition CT scans were obtained, with the carina and septum used as cranial and caudal landmarks, respectively. For the analysis of adipose tissue radiomic features, the original DICOM (Digital Imaging and Communications in Medicine) images were transferred to a dedicated workstation and further post-processed using 3D Slicer (see below).

[0332] Radiomic feature extraction of peri-atrial adipose tissue

[0333] Calculation of radiomic features in peri-atrial fat tissue in CT scans was performed using 3D Slicer software (v.4.9.0-2017-12-18r26813, available at http: / / www.slicer.org; see Fedorov, A. et al., 3D Slicer as an image computing platform for quantitative imaging networks, Magnetic Resonance Imaging 30, 1323-1341, doi: 10.1016 / j.mri.2012.05.001 (2012)). Segmentation of fat tissue was performed by locating a manual outline of peri-atrial epicardial fat tissue between the levels of the pulmonary veins and using a -190 to -30 Hounsfield unit (HU) mask for fat identification. Subsequently, the segmented adipose tissue was used to calculate and extract a series of radiomic features using the SlicerRadiomics extension of 3D Slicer, which incorporates the Pyradiomics library of radiomic features into 3D Slicer (see van Griethuysen, JJM et al., A computational radiomics system for decoding radiomic phenotypes, Cancer Research 77, e104-e107, doi: 10.1158 / 0008-5472.CAN-17-0339 (2017)). The original HU values ​​of the segmented adipose tissue were used to calculate shape-related radiomic features and first-order radiomic features. In order to calculate texture features (GLCM, GLDM, GLRLM, GLSZM, and NGTDM), the adipose tissue voxels were discretized into 16 equal-width bins (width of 10 HU) to reduce noise while allowing sufficient resolution to detect biologically important spatial changes in adipose tissue attenuation. As described above, first-order statistics and texture-based statistics were also calculated for the three-dimensional wavelet transform of the raw image data, resulting in eight additional radiomic feature sets.

[0334] Collection of human myocardial samples

[0335] During CABG, myocardial tissue samples were collected from the right atrial appendage (as previously described in Antonopoulos, A.S. et al., Reciprocal Regulation of Epicardial Adipose Tissue and Myocardial Redox Status by PPAR-γ / Adiponectin Signaling, Circulation Research 118, 842-855, doi:10.1161 / CIRCRESAHA.115.307856 (2016)) and transferred to the laboratory and placed in an oxygenated (95% O 2 / 5%CO 2 ) in ice-cold buffer for further studies.

[0336] Myocardial superoxide measurement

[0337] Measurement of myocardial O in right atrial appendage samples using lucigenin (5 μmol / L) enhanced chemiluminescence 2 Products, as described in Antoniades, C. et al., Myocardial redox state predicts in-hospital clinical outcome after cardiac surgery effects of short-term pre-operative statin treatment, J Am Coll Cardiol 59, 60-70, doi: 10.1016 / j.jacc.2011.08.062 (2012). Myocardial tissue was homogenized in ice-cold Krebs HEPES buffer pH 7.35 in the presence of protease inhibitors using a pre-cooled Polytron homogenizer (Roche Applied Science, Indianapolis, IN).

[0338] Study on gene expression in human atrial myocardium

[0339] Atrial myocardial tissue samples were snap frozen in QIAzol (Qiagen) and stored at -80° C. RNA was extracted using RNeasy Micro or Mini kits (Qiagen) and ribonucleic acid was converted to complementary DNA (Quantitect Rev. Transcription Kit - Qiagen). The cDNA was then subjected to qPCR using TaqMan probes (Applied Biosystems) for: TNFA (assay ID Hs01113624_g1), IL6 (assay ID Hs00985639_m1), IFNG (assay ID Hs00989291_m1), COL1A1 (assay ID Hs00164004_m1), NPPA (assay ID Hs00383230_g1), BNP (assay ID Hs00173590_m1) and PGK1 was used as a housekeeping gene (assay number Hs00943178_g1) Reactions were performed in triplicate in 384-well plates on an ABI 7900HT Fast Real-Time PCR System (Applied Biosystems), using 5 ng of cDNA per reaction. The efficiency of the reaction in each plate was determined based on the slope of the standard curve; the expression of each gene of interest relative to its housekeeping gene was calculated using the Pfaffl method.

[0340] Statistical analysis

[0341] Principal component analysis: In cohort A, all 843 computed adipose tissue radiomic signatures were included in principal component analysis to identify principal components that described the majority of phenotypic variation in the study population. A scree plot of the principal components against their eigenvalues ​​was constructed. All principal components with eigenvalues ​​greater than 1 were then included in a logistic regression model with backward elimination of AF classification as the dependent variable to identify principal components (PCs) of peri-atrial fat radiomic signatures that were independently associated with AF.

[0342] Unsupervised clustering of the study population by adipose tissue radiomic signatures: Initially, the association of each of the radiomic signatures in group A with AF was explored by receiver operating characteristic (ROC) curve analysis for AF classification. The statistical strength of association (p-value, where p-value is a probability value or asymptotic significance) of all radiomic signatures was graphically represented on Manhattan plots and further stringently adjusted by Benjamini-Hochberg correction using a false discovery rate of 0.10. The final 33 radiomic signatures that were independently associated with AF were then used to hierarchically cluster the population of group A (using the Ward D2 method and squared Euclidean distance, hclust R package, e.g., as in Langfelder, P. and Horvath, S., Fast R Functions for Robust Correlations and Hierarchical Clustering Journal of Statistical Software 46 (2012)). The changes in each of the selected radiomic features in the observations across the Group A cohort were presented as correlation heatmaps with row dendrograms representing the clustering of patients. Differences in the prevalence of AF, risk factors, or distribution of biochemical data between the two patient clusters were then explored using chi-square tests or independent t tests, as appropriate.

[0343] Feature selection and machine learning for the radiomic fingerprint of AF on peri-atrial fat: The relationships between selected radiomic features were visually inspected by correlation plots. To further perform feature selection and develop a radiomic score for peri-atrial fat, the cohort was divided into a training (80%) cohort and a testing (20%) cohort using a random seed. The features were then fed into a machine learning algorithm (caret R package, see Kuhn, M., Caret: Classification and regression training Vol. 1 (2013)) to identify the features with the best performance (AUC or c statistics) for AF classification (i.e., distinguishing people with AF from those without AF). The model was trained using 5-fold internal cross-validation (5×3 times) repeated 3 times. The accuracy of the final model was evaluated in the training cohort and then externally validated in the testing cohort. The final model was used to form a radiomic score for peri-atrial fat based on the predicted probability of AF.

[0344] Validation of the radiomic score of peri-atrial fat for atrial tissue phenotype: The developed radiomic score of peri-atrial fat was externally tested in an independent cohort of patients undergoing CABG (n=225, Group B) to validate it for atrial tissue phenotype. A similar approach was used to extract the radiomic signature of peri-atrial fat in Group B patients and then assigned to each of the patients based on the developed Group A algorithm. The association of the peri-atrial fat radiomic score for atrial gene expression profile, atrial redox status was assessed in bivariate analysis using unpaired t-tests between the two groups, while the risk of postoperative AF was explored by Kaplan-Meier curves in survival analysis as appropriate.

[0345] Continuous variables between the two groups were compared by Student's t-test, while categorical variables were compared by Pearson's Chi-square test. Analyses were performed using R v3.4 (packages: aret, hclust) and SPSS version 25.0. Unless otherwise stated, all tests were two-sided and α was set to 0.05.

[0346] result

[0347] Radiomics feature extraction and principal component analysis

[0348] Study design Figure 1In cohort A, a total of 2,246 patients underwent diagnostic coronary CT angiography (Cleveland Clinic, USA). The CT imaging dataset was used to extract the radiomic features of peri-atrial fat. A nested case-control analysis (n=310) of 155 subjects with AF versus 155 control individuals (sinus rhythm) was performed with full 1:1 matching of subjects on age, sex, cardiovascular risk factors, and scan acquisition details (Table 4).

[0349] Table 4. Demographic characteristics of the study population

[0350]

[0351] ACEi: angiotensin-converting enzyme inhibitor; ARB: angiotensin receptor blocker; HDL: high-density lipoprotein; ICD: intracardiac defibrillator; LDL: low-density lipoprotein; PPM: permanent pacemaker.

[0352] A total of 843 radiomic features were calculated by segmenting the peri-atrial fat tissue (surrounding the left atrium at the level located between the pulmonary veins) as summarized in Table 5. These included 15 shape-related features, 18 first-order statistics, 15 gray-level co-occurrence matrix (GLCM) features, 18 gray-level dependence matrix (GLDM) features, 16 gray-level run-length matrix (GLRLM) features, 16 gray-level size zone matrix (GLSZM) features, and 5 neighboring gray-level tone difference matrix (NGTDM) features (as defined in Tables R1 to R7), as well as eight wavelet transforms for each of these features.

[0353] Table 5: Breakdown of radiomic features.

[0354]

[0355] Initially, exploratory data analysis was performed by reducing the raw radiomics dataset of potentially relevant features to its principal components. A total of 86 components accounted for 99.5% of the variation in the study population (scree plot, Figure 2a ), while the first three components explain 56% of the observed variation ( Figure 2b Among the 51 independent components with eigenvalues ​​> 1, five (PCs 4, 13, 16, 22, 26, and 28) were significantly independently associated with AF in logistic regression ( Figure 2c ), thus indicating that texture-related features of peri-atrial fat on standard CT images contain rich extractable information that is clearly associated with AF and also with possible atrial tissue phenotype.

[0356] Unsupervised clustering based on radiomic phenotyping of peri-atrial adipose tissue

[0357] Since principal components are intrinsic to the sample population studied and not transferable as quantifiable biomarkers, the peri-atrial radiomic features themselves were analyzed. The intercorrelations between the 843 radiomic features of peri-atrial fat are shown in Figure 2. Figure 2d From the initial pool of 843 quantitative radiomics features, a set of 33 features was significantly associated with AF after rigorous statistical adjustment (Manhattan plot, Figure 2e ). Unsupervised hierarchical clustering of group A using the radiomic features of peri-atrial adipose tissue identified two distinct patient clusters that differed significantly in the incidence of AF (heat map, Figure 3a These findings support that the presence of AF is associated with certain radiomic fingerprints of epicardial regions, such as peri-atrial fat areas, which could be useful for extracting imaging biomarkers of atrial disease.

[0358] Machine learning to identify radiomic signatures of atrial fibrillation on peri-atrial fat

[0359] After demonstrating proof of concept that the radiomic signature of peri-atrial adipose tissue is different in the presence of AF, a radiomic signature or “fingerprint” of AF on peri-atrial adipose tissue was constructed. Figure 3b The inter-correlations and hierarchical clustering of 33 selected radiomic features of peri-atrial fat are presented in Figure 2. The correlation plots showed that the radiomic features of peri-atrial fat tissue were clustered into different groups of related features (one shape-related group and five texture-related groups).

[0360] To identify selected radiomic features that can be used to construct a radiomic fingerprint of AF on peri-atrial fat, the cohort of Group A was split into a training dataset (80%) and a testing dataset (20%) for model training and testing, respectively, using a random seed. First, recursive feature elimination using a random forest algorithm was used to: a) find the number of features required to maximize the accuracy of the model with AF; and b) select the best features to be included in the model. Of the 33 selected radiomic features, a set of 15 features maximized the algorithm's diagnostic accuracy for AF ( Figure 4a ).

[0361] These 15 radiomic features of peri-atrial fat (Table 2) were then fed as independent variables into various machine learning algorithms, and the performance of each algorithm for AF classification was explored. After three repetitions of internal 5-fold cross validation (5 × 3 folds), the algorithm with the best performance for AF classification was identified ( Figure 4b, c). The extreme gradient boosting algorithm (xgbDART, originally described in VK Rashmi and R. Gilad-Bachrach, DART: Dropouts meet Multiple Additive Regression Trees, Journal of Machine Learning Research (JMLR) (2015), and available from https: / / xgboost.readthedocs.io / ) used 15 radiomic features and achieved an accuracy of 74.5% in classifying AF (distinguishing between people with AF and those without AF) in the training dataset (Figure 4d). The performance of the same algorithm was then evaluated in the test dataset, where 70% of the cases were correctly classified (Figure 4e). The calculated value of the radiomic label is P×10, where P is the predicted probability of the presence of AF (P) output by the decision tree of the radiomic label.

[0362] The parameters used for the xgbDART algorithm are as follows: max_depth = 2, eta = 0.4, rate_drop = 0.5, skip_drop = 0.05, subsample = 0.5, colsample_bytree = 0.8, nrounds = 150. These parameters were optimized by maximizing the ROC value for AF discrimination in group 1. Other parameters were kept at their default values, specifically the tuning parameter "γ" was kept at 0, and the tuning parameter "min_child_weight" was kept at 1.

[0363] Study whether radiomic signatures detect myocardial disease (myocardial tissue redox state, fibrosis, and inflammation)

[0364] Having identified the fingerprint of AF on peri-atrial fat, we investigated whether this radiomic signature could be used to assess changes in atrial biology. In the Group B cohort, peri-atrial fat was imaged by CT in 225 patients undergoing coronary artery bypass grafting (AF prevalence = 7.3%), and atrial tissue samples were collected perioperatively for gene expression studies and to assess myocardial redox status. For each patient, peri-atrial fat tissue texture was assessed by the same radiomic score ( Figure 5a The peri-atrial fat radiomics score was positively correlated with the gene expression levels of COL1A1, natriuretic peptides (ANP, BNP), and pro-inflammatory genes (such as IL6 and TNFA) in the atrial myocardium ( Figure 5b , c). Higher radiomic scores of peri-atrial fat were also associated with increased superoxide production in human atrial tissue ( Figure 5d ).

[0365] Finally, this developed radiomic score of peri-atrial fat was investigated for its diagnostic value for AF and atrial tissue phenotype and also for predicting the risk of postoperative AF. A cut-off value of 6 for the radiomic score of peri-atrial fat increased the risk of postoperative AF ( Figure 5e , f). In the survival analysis, the peri-atrial fat tissue radiomics score was significantly associated with postoperative AF ( Figure 5g ).

[0366] Validation of the surrogate radiomics signature of the invention

[0367] The above discussion shows that the radiomic signature calculated based on the 15 radiomic features identified using an unbiased machine learning approach and listed in Table 2 significantly improves the discriminatory value of the model for arrhythmias, and therefore potential cardiac health, for which arrhythmias are used as surrogate markers. Therefore, the radiomic signature of the present invention is able to distinguish potential pathologies such as myocardial fibrosis, oxidative stress (i.e., redox state), and inflammation. In order to verify the usefulness of the alternative radiomic signatures of the present invention comprising different selections of radiomic features, a series of several different radiomic signatures were tested for AF detection, again using the AF as a surrogate marker for potential myocardial health. The results are shown in Table 6.

[0368] The radiomic signatures of Example 1 were calculated based on 15 radiomic features identified using an unbiased machine learning approach and are listed in Table 2. In Example 2, each of the 15 original radiomic features has been replaced by the most collinear radiomic feature, and in Example 3, each of the 15 original radiomic features has been replaced by the least collinear radiomic feature (see Table 3). Finally, in Example 4, each of the original radiomic features replaced a different radiomic feature from the same cluster (clusters AD in Table 1). The accuracy of each of these signatures was tested to detect atrial fibrillation (AF) in Group 1, and the results are shown in Table 6.

[0369] Table 6. Performance of various radiomics features including different combinations of radiomics features

[0370]

[0371]

[0372] As can be clearly seen from Table 6, all tested radiomic signatures of the present invention provide high accuracy for the detection of atrial fibrillation. Thus, the data presented in Table 6 demonstrate that, regardless of which features are selected from each of the identified clusters or groups, the radiomic signatures of the present invention provide an improved prediction of cardiovascular risk compared to previously used models. In addition, Figure 4a The data presented in suggest that a radiomic signature can be calculated using fewer than 15 radiomic features and that the radiomic signature will still be able to distinguish cardiac health. Figure 4a It can be seen that a radiomic signature consisting of as few as two radiomic features has useful discriminatory value for cardiac health. Figure 4a It can be seen that when at least three radiomic features are included in the signature, the accuracy of the signature increases dramatically. Therefore, the radiomic signature preferably includes at least three radiomic features, and wherever at least two radiomic features are mentioned herein, it should be understood that this may mean at least three radiomic features.

[0373] Example 2

[0374] Deep learning methods for automated CT scan segmentation

[0375] 400 diagnostic coronary CT angiograms were manually segmented to identify three discrete segments: the epicardial and atrial appendage regions of interest ( Figure 6a , Note A), the epicardial region of the anterior wall of the left atrium ( Figure 6a , Annotation B) and the epicardial region of the left atrial appendage of interest ( Figure 6a , Annotation C). The entire left atrium (LA) itself is also segmented ( Figure 6a , Note D). The scan was segmented using a wider HU range (-190HU to +150HU).

[0376] These scans and the manually segmented regions of interest were fed into a deep learning system that was able to learn the features of the manually segmented anatomical volumes. The deep learning approach employed allowed the computation of a multi-layer neural network and used a convolutional neural network (CNN). The results of the deep learning automatic segmentation from the first 300 scans (training set) were very impressive ( Figure 6b ), where the agreement between human-segmented scans and machine-segmented scans in the external dataset was 100% consistent across 100 scans from the external validation set ( Figure 6c ). This automated segmentation can be combined with radiomic feature extraction to provide a complete application for automated radiomic analysis of left atrial periventricular tissue.

[0377] Example 3

[0378] The following examples are generally similar to Example 1 unless otherwise noted.

[0379] To demonstrate that stroke is associated with distinct LA peri-tissue radiomic signatures, an independent cohort of 98 patients undergoing diagnostic coronary CT angiography (Erlangen, Germany) was analyzed. Patients with a history of stroke (n = 49) were matched 1:1 with control subjects without a known history of stroke (n = 49). Both groups were matched for age, sex, cardiovascular risk factors, and scan acquisition details.

[0380] This study was designed to identify radiomic features of epicardial regions independently associated with stroke to develop a correlated radiomic score of left atrial periventricular tissue associated with cardiac and specifically atrial health. The cohort sample was taken from a mixed population of patients with atrial fibrillation (AF) with a 5-year follow-up stroke rate of 2.4%.

[0381] Scans were segmented manually by a single reader at the Oxford Academic Cardiovascular Computed Tomography Core Lab. For each scan, the four broadly attenuation-defined segments described (-190 HU to +150 HU) were identified. These segments included 1) the atrial septum, 2) the anterior surface of the LA, and 3) the left atrial appendage ( Figure 6a ). Radiomic features were extracted from each segment as well as the combined segments. Variables were first extracted only for the LA peri-fat tissue (-190HU to -30HU) and then for the entire voxel range within each segment (-190HU to +150HU).

[0382] To construct the most highly predictive radiomic signature of atrial myopathy leading to stroke in peria-LA tissue, machine learning segments were applied to each of three independent segments (1-3), three merged segments of adjacent independent segments (1+2, 2+3, and 1+2+3), and both the narrower peria-LA adipose tissue and the wider peria-LA Hounsfield extent. There were 12 different analysis groups.

[0383] For each analysis group, 98 patients were split into a training (80%) dataset and an external validation (20%) dataset using random seeds. First, recursive feature elimination using a random forest algorithm was used to: a) find the number of features required to maximize the accuracy of the model with stroke; and b) select the best features for the model.

[0384] A set of 16 radiomic features was found to maximize the algorithm's diagnostic accuracy for stroke. These 16 features are detailed in Table 2b. The 16 radiomic features were calculated for both the narrower LA peri-LA fat tissue and the wider LA peri-LA tissue HU range, and then the wider LA peri-LA tissue HU range was fed into various machine learning algorithms as an independent variable, and the performance of each algorithm in classifying participants with stroke was explored. After 3 repetitions of internal 5-fold cross validation ( Figure 6d ), derived the Atriomic Stroke algorithm (epicardial radiomic signature predicts stroke) for optimal stroke prediction. The most successful prediction model used 16 radiomic features calculated for combined segments of the atrial septum and the anterior LA wall. The Atriomic Stroke algorithm employed an extreme gradient boosting algorithm using 16 radiomic features that had extremely high accuracy (0.87±0.04) for stroke classification in an external validation dataset ( Figure 6e - Row A).

[0385] In more detail, principal component analysis was used to detect principal components of peri-atrial tissue radiomics signatures independently associated with stroke occurrence using an eigenvalue greater than 1 as the cutoff value for components included in a logistic regression model with backward elimination for stroke classification as the dependent variable.

[0386] The Atriomic Stroke Algorithm was developed using the same machine learning approach as outlined in Example 1, employing the extreme gradient boosting package in the R environment to achieve a model with the best performance, as measured by AUC for stroke classification (i.e., distinguishing those who went on to have a stroke from those who did not). The model was trained using a 5-fold internal cross validation (5×3 times) repeated 3 times. The accuracy of the final model was evaluated in the training cohort and then externally validated in the test cohort. The final model was used to develop the Atriomic Stroke Algorithm based on the predicted stroke probability in this sample.

[0387] Atriomic stroke algorithm identifies genes associated with inflammation and fibrosis within atrial tissue

[0388] An independent cohort of 86 patients undergoing coronary artery bypass grafting (CABG) was assembled. Patients underwent coronary CT angiography scans, and perioperative atrial tissue samples were collected for targeted myocardial tissue phenotyping (as previously described). This cohort was used for external validation of the developed Atriomic stroke algorithm for atrial biology.

[0389] The association of peri-atrial tissue radiomics scores with atrial gene expression profiles was assessed in bivariate analysis using unpaired t-tests between groups. The Atriomic Stroke Algorithm was associated with adverse atrial health processes and could detect high atrial expression of collagen (COL1A1) reflecting high fibrotic activity and high atrial expression of pro-inflammatory genes including TNFA reflecting active inflammation (see Figure 6f ).

[0390] Atriomic stroke algorithm outperforms current stroke risk prediction tools

[0391] To demonstrate that the Atriomic Stroke Algorithm can be used clinically, an independent cohort of 98 patients undergoing diagnostic coronary CT angiography (used for algorithm development) was used. 2 DS 2 -The ability of VASc score in predicting stroke. This is a widely used clinical scoring system for stroke stratification in clinical practice. It was demonstrated that the Atriomic Stroke Algorithm is far superior to CHA in stroke prediction on CRISP-CT 2 DS 2 -VASc, where the change in AUC was 0.12 (p=0.005) (see Figure 6e ).

[0392] Validation of the surrogate radiomics signature of the invention

[0393] The above discussion shows that the radiomics signature calculated based on the 16 radiomics features identified using an unbiased machine learning method and listed in Table 2b significantly improves the model's discriminatory value for ischemic stroke, and therefore potential heart health, for which ischemia is used as a surrogate marker. Therefore, the radiomics signature of the present invention is able to distinguish potential conditions such as myocardial fibrosis, oxidative stress (i.e., redox state) and inflammation. In order to verify the usefulness of the alternative radiomics signatures of the present invention comprising different selections of radiomics features, a series of several different radiomics signatures were tested for stroke detection. The results are shown in Table 7.

[0394] The radiomics label of example 1 is based on 16 radiomics features calculated using unbiased machine learning method identification, and is listed in table 2b.In example 2, each radiomics feature in 16 original radiomics features has been replaced by the radiomics feature with its most collinearity, and in example 3, each original radiomics feature in 16 original radiomics features has been replaced by the radiomics feature with its least collinearity (referring to table 3b).Finally, in example 4, each original radiomics feature in radiomics features replaces the different radiomics features (clustering AD in table 1b) from the same cluster.The accuracy of each label in these labels has been tested, to detect the stroke in the CRISP-CT queue, and the results are shown in table 7.

[0395] Table 7. Performance of various radiomic signatures including different combinations of radiomic features

[0396]

[0397]

[0398] It is clear from Table 7 that all tested radiomic signatures of the present invention provide high accuracy for ischemic stroke detection. Thus, the data presented in Table 7 demonstrate that, regardless of which features are selected from each of the identified clusters or groups, the radiomic signatures of the present invention provide improved prediction of cardiovascular risk compared to previously used models.

[0399] Summary of findings

[0400] The studies outlined above demonstrate that radiomic phenotyping of epicardial regions of the human heart (e.g., peri-atrial tissue) can be used to assess phenotypic changes associated with cardiac health. Following an unbiased feature selection process and machine learning for algorithm training, internal cross-validation, and external testing, it is possible to identify a radiomic fingerprint of myocardial health (e.g., using cardiac conditions such as arrhythmias as surrogate markers of myocardial health) in epicardial tissue and develop a radiomic signature or score to characterize the epicardial region and, therefore, also indirectly the adjacent myocardium.

[0401] The ECR radiomics signature of the invention also adds incremental value beyond traditional risk factors in predicting the development of cardiac conditions such as arrhythmias, e.g., post-operative atrial fibrillation, and captures myocardial biology, specifically atrial biology, such as fibrosis and myocardial oxidative stress.

[0402] Surprisingly, a radiomic signature need not be constructed from the radiomic features that are most independently closely associated with cardiac conditions or myocardial disease. Rather, it may actually be advantageous to include radiomic features selected from different "clusters" of related or similar radiomic features, rather than only those radiomic features that are most independently associated with cardiac conditions or myocardial disease. Furthermore, important radiomic features may be replaced with collinear equivalents while still providing an effective signature indicative of myocardial disease.

[0403] A particularly attractive aspect of the present invention is that it can be performed based on historical medical imaging data that has been previously collected. The signatures of the present invention can be derived and calculated based on historical imaging data, and thus the present invention provides a convenient tool for evaluating a large number of patients without having to perform additional scans. Therefore, the method of the present invention does not have to include a step of collecting medical imaging data, and can be performed based on a post hoc analysis of existing medical imaging data.

Claims

1. A method for characterizing an epicardial region comprising epicardial tissue, the method comprising calculating a value of an epicardial radiomics signature of the epicardial region using medical imaging data; wherein the epicardial radiomic signature is calculated based on measured values ​​of a plurality of epicardial radiomic features of the epicardial region, the measured values ​​of the epicardial radiomic features being calculated based on the medical imaging data; as well as Wherein the epicardial radiomic signature provides a measure of the texture of the epicardial tissue.

2. The method according to claim 1, wherein: The epicardial radiomic signature is indicative of cardiac health.

3. The method according to claim 1 or 2, wherein: The epicardial radiomic signature predicts the likelihood of a subject suffering from a cardiac condition.

4. The method according to claim 3, wherein: The cardiac condition is arrhythmia.

5. The method of claim 1 or 2, wherein the epicardial radiomic signature predicts the likelihood that a subject will experience a stroke.

6. The method according to claim 5, wherein: The epicardial region includes the area surrounding the atria.

7. The method according to claim 6, wherein: The peri-atrial region includes the atrial septal epicardial region and the left atrium or right atrium anterior epicardial region.

8. The method according to claim 1 or 2, wherein: The plurality of epicardial radiomics signatures include at least two epicardial radiomics signatures selected from groups 1 to 15 of epicardial radiomics signatures, wherein the at least two epicardial radiomics signatures are each selected from a different group, and wherein: Group 1 consists of the inverse difference moment HHH, the normalized inverse difference HHH, the contrast HHH for the gray-level co-occurrence matrix, the range HHH, the complexity HHH, the maximum value HHH, the large dependence high gray-level advantage HHH, and the short run low gray-level advantage HHH; Group 2 consists of minimum value LHH, long run low grayscale advantage LHH, short run high grayscale advantage LHH, high grayscale advantage LHH, high grayscale run advantage LHH, small area high grayscale advantage LHH, high grayscale region advantage LHH, autocorrelation LHH, joint average LHH, average sum LHH, short run low grayscale advantage LHH, long run high grayscale advantage LHH, range LHH, low grayscale advantage LHH, low grayscale run advantage LHH, large dependence low grayscale advantage LHH, low grayscale region advantage LHH, small area low grayscale advantage LHH, small dependence high grayscale advantage LHH, complexity LHH, large dependence high grayscale advantage LHH, cluster prominence LHH, grayscale variance LHH for grayscale size region matrix and maximum value LHH; Group 3 consists of low grayscale area advantage LLL, short run low grayscale advantage LLL, low grayscale run advantage LLL, low grayscale advantage LLL, long run low grayscale advantage LLL, small area low grayscale advantage LLL, small dependence low grayscale advantage LLL, large area low grayscale advantage LLL and large dependence low grayscale advantage LLL; Group 4 consists of maximum probability LLL, joint energy LLL, joint entropy LLL, maximum probability, joint energy, joint entropy, normalized grayscale unevenness, energy LHL, uniformity, size area unevenness, sum entropy, normalized grayscale unevenness, entropy, normalized grayscale unevenness LLL, uniformity LLL, mean, normalized grayscale unevenness LLL, root mean square, interquartile range, sum entropy LLL, robust mean absolute deviation, size area unevenness HLL, size area unevenness LHL, 10th percentile, energy HHL, median, dependent unevenness LHL, entropy LLL, mean absolute deviation, energy LLH, run entropy LLL, interquartile range LLL, size area unevenness LLH, energy HLL, sum of squares, dependent unevenness, robust mean absolute deviation LLL, 10th percentile LLL, energy LHH, dependent unevenness, and run entropy; Group 5 consisted of busyness LHH, intensity LHH, intensity HHH, busyness HHH, busyness LHL, and large area low grayscale dominance LHH; Group 6 consists of regional entropy LLL, dependence entropy LLL, root mean square LLL, mean LLL, run entropy, dependence entropy, median LLL, median, mean, 10th percentile LLL, uniformity, normalized grayscale non-uniformity for grayscale dependence matrix, root mean square, 90th percentile, entropy, 10th percentile, interquartile range LLL, run entropy LLL, robust mean absolute deviation LLL, normalized grayscale non-uniformity for grayscale dependence matrix LLL and uniformity LLL; Group 7 consists of run entropy LLL, entropy LLL, mean absolute deviation LLL, mean absolute deviation, robust mean absolute deviation, robust mean absolute deviation LLL, variance, variance LLL, grayscale variance for grayscale size region matrix LLL, grayscale variance for grayscale size region matrix, interquartile range, interquartile range LLL, entropy, grayscale variance for grayscale dependence matrix LLL, root mean square, run entropy, grayscale variance for grayscale dependence matrix, sum entropy, sum of squares, sum entropy LLL, sum of squares LLL, cluster tendency, cluster tendency LLL, joint entropy, root mean square LLL, contrast for graylevel co-occurrence matrix, joint entropy LLL, cluster Class prominence, cluster prominence LLL, low grayscale advantage, short run low grayscale advantage, low grayscale run advantage, long run low grayscale advantage, regional entropy LLL, low grayscale regional advantage, uniformity LLL, normalized grayscale unevenness LLL for grayscale dependency matrix, normalized grayscale unevenness LLL for grayscale size region matrix, 10th percentile, 10th percentile LLL, normalized grayscale unevenness for grayscale dependency matrix, normalized grayscale unevenness for grayscale size region matrix, mean, uniformity, joint energy, median, joint energy LLL, maximum probability, mean LLL and maximum probability LLL; Group 8 consisted of the maximum 3D diameter, the maximum 2D diameter slice, the maximum 2D diameter column, and the long axis; Group 9 consists of long run high grayscale advantage LLL, high grayscale advantage LLL, high grayscale run advantage LLL, high grayscale area advantage LLL, short run high grayscale advantage LLL, autocorrelation LLL, joint average LLL, small area high grayscale advantage LLL, small dependence high grayscale advantage LLL, minimum value LLL, large dependence low grayscale advantage LLL, large dependence high grayscale advantage LLL, large area high grayscale advantage LLL, large area low grayscale advantage LLL, long run low grayscale advantage LLL, low grayscale advantage LLL, low grayscale run advantage LLL and short run low grayscale advantage LLL; Group 10 consists of the joint mean LLL, autocorrelation LLL, long run high grayscale advantage LLL, high grayscale advantage LLL, high grayscale run advantage LLL, high grayscale area advantage LLL, short run high grayscale advantage LLL, small area high grayscale advantage LLL, small dependence high grayscale advantage LLL, minimum value LLL, large dependence low grayscale advantage LLL, large dependence high grayscale advantage LLL, large area high grayscale advantage LLL, large area low grayscale advantage LLL, long run low grayscale advantage LLL, low grayscale advantage LLL, low grayscale run advantage LLL, short run low grayscale advantage LLL and low grayscale area advantage LLL; Group 11 consists of difference entropy LLL, difference mean LLL, contrast LLL for adjacent gray tone difference matrix, difference entropy, inverse difference LLL, contrast for gray level co-occurrence matrix, difference variance, inverse difference moment LLL, difference mean, inverse variance LLL, inverse variance, difference variance LLL, inverse difference, inverse difference moment, normalized inverse difference moment, normalized inverse difference, contrast for adjacent gray tone difference matrix, joint entropy, sum entropy LHL, joint energy LHL, run entropy LHL, normalized size area unevenness LLL, small area advantage LLL, short run advantage, normalized size area unevenness, small area advantage, normalized gray level unevenness LHL for gray level size area matrix, joint entropy LHL, short run advantage LLL , Small Dependence Advantage LLL, Normalized Dependence Non-Uniformity LLL, Normalized Grayscale Non-Uniformity LHL for Grayscale Dependence Matrix, Small Dependence Advantage, Entropy LHL, Long Run Advantage LLL, Mean Absolute Deviation LHL, Robust Mean Absolute Deviation LHL, Uniformity LHL, Interquartile Range LHL, Joint Energy, Normalized Run Length Non-Uniformity LLL, Run Percentage LLL, Region Percentage LLL, Long Run Advantage, Sum of Squares LHL, Complexity LLL, Normalized Run Length Non-Uniformity, Run Percentage, Region Percentage, Clustering Propensity LHL, Run Variance LLL, Large Dependence Advantage LLL, Normalized Dependence Non-Uniformity, Run Variance, Grayscale Variance LHL for Grayscale Dependence Matrix, Large Area advantage LLL, variance LHL, grayscale variance LHL for grayscale size area matrix, large dependence advantage, large area advantage, maximum probability LHL, root mean square LHL, difference entropy LHL, grayscale variance LHL for grayscale run length matrix, area variance LLL, dependence variance LLL, inverse difference LHL, inverse difference moment LHL, area variance, large area high grayscale advantage, 90th percentile LHL, sum entropy LLH, difference mean LHL, sum of squares, dependence entropy LHH, contrast LHL for grayscale co-occurrence matrix, joint energy HLL, difference entropy HLL, difference variance LHL, dependence variance, maximum probability HLL, complexity, joint entropy HLL, joint energy LLL, sum entropy LHH, inverse Variance LHL, 90th percentile LLH, inverse difference HLL, inverse difference moment HLL, difference variance HLL, clustering tendency LHH, difference mean HLL, clustering tendency LLH, contrast HLL for gray-level co-occurrence matrix, run entropy LHH, inverse variance HLL, joint energy LLH, joint energy HHL, joint entropy LLL, run entropy LLH, joint entropy LLH, large dependence high gray-level advantage, maximum probability HHL, joint entropy HHL, sum entropy HHL, normalized gray-level unevenness HLL for gray-level dependency matrix, robust mean absolute deviation LLH, uniformity HLL, cluster prominence LHL, complexity LHL, entropy LLH, normalized gray-level unevenness LLH for gray-level dependency matrix,Mean absolute deviation LLH, run entropy HHL, uniformity LLH, normalized grayscale unevenness LLH for grayscale size region matrix, interquartile range HLL, interquartile range LLH, maximum probability LLH, robust mean absolute deviation HLL, normalized grayscale unevenness HHL for grayscale dependency matrix, long run advantage LHL, robust mean absolute deviation HHL, run variance LHL, uniformity HHL, interquartile range HHL, joint entropy LHH, sum of squares LLH, 10th percentile HHL, 90th percentile HHL, entropy HHL, clustering tendency HHL, normalized grayscale unevenness HLL for grayscale size region matrix, flat Mean absolute deviation HHL, 10th percentile LHL, difference entropy HHL, sum of squares HHL, contrast LLL for gray-level co-occurrence matrix, gray-level variance HHL for gray-level dependence matrix, variance HHL, entropy HLL, normalized gray-level unevenness HHL for gray-level size region matrix, gray-level variance HHL for gray-level size region matrix, inverse difference HHL, joint energy LHH, root mean square HHL, short run advantage LHL, sum of squares LHH, 10th percentile LHH, inverse difference moment HHL, mean absolute deviation LHH, run percentage LHL, region percentage LHL, normalized dependence unevenness LHL, entropy LHH, for Normalized grayscale unevenness LHH for grayscale size region matrix, large dependence advantage LHL, interquartile range LHH, maximum probability LHH, small dependence advantage LHL, uniformity LHH, large area advantage LHL, robust mean absolute deviation LHH, root mean square LLH, difference mean HHL, small dependence low grayscale advantage, entropy LLL, grayscale variance for grayscale dependency matrix, normalized run unevenness LHL, variance, region variance LHL, cluster prominence HHL, dependency variance LHL, grayscale variance LHH for grayscale dependency matrix, root mean square LHH, grayscale variance for grayscale size region matrix, variance LHH, grayscale co-occurrence matrix The matrix consists of contrast HHL, dependence entropy HHL, normalized size area unevenness LHL, small area advantage LHL, sum entropy HHH, difference variance HHL, grayscale variance HHL for grayscale run length matrix, dependence entropy LHL, grayscale variance LLH for grayscale size area matrix, run entropy HLL, variance LLH, 90th percentile LHH, mean absolute deviation HLL, grayscale variance LLH for grayscale dependency matrix, clustering tendency HHH, difference entropy LLH, inverse difference moment LLH, normalized grayscale unevenness LHH for grayscale dependency matrix, inverse difference LLH, mean absolute deviation and 90th percentile HLL; Group 12 consists of sum entropy HHH, cluster tendency HHH, cluster prominence HHH, joint entropy HHH, joint energy HHH, difference entropy HHH, difference variance HHH, sum of squares HHH, normalized grayscale unevenness HHH for grayscale size region matrix, uniformity HHH, entropy HHH, grayscale variance HHH for grayscale dependency matrix, grayscale variance HHH for grayscale size region matrix, root mean square HHH, variance HHH, mean absolute deviation HHH, 10th percentile HHH, robust mean absolute deviation HHH, 90th percentile HHH, interquartile range HHH, normalized grayscale unevenness HHH for grayscale dependency matrix, grayscale variance HHH for grayscale run length matrix, and Entropy LHH, joint entropy HHL, difference entropy HHL, clustering tendency LHH, joint energy HHL, long run advantage HHL, maximum probability HHL, short run advantage HHL, large area low gray level advantage HHL, sum entropy HHL, normalized size area unevenness HHL, sum of squares HHL, difference variance HHL, joint entropy LHH, normalized run length unevenness HHL, clustering tendency HHL, contrast HHL for gray level co-occurrence matrix, difference mean HHL, inverse difference HHL, large dependence advantage HHL, run percentage HHL, run variance HHL, small area advantage HHL, inverse difference moment HHL, small dependence advantage HHL, sum of squares LHH, gray level variance LH for gray level dependence matrix H, Root mean square HHL, Variance HHL, Difference variance LHH, Entropy HHL, Grayscale variance HHL for grayscale dependency matrix, Grayscale variance HHL for grayscale size region matrix, Grayscale variance LHH for grayscale size region matrix, Mean absolute deviation HHL, Root mean square LHH, Variance LHH, Joint energy HLH, 90th percentile HHL, Joint energy LHH, Normalized dependency unevenness HHL, Entropy LHH, Normalized grayscale unevenness HHL for grayscale size region matrix, Joint entropy HLH, Uniformity HHL, Cluster prominence HHL, Cluster prominence LHH, Mean absolute deviation LHH, 10th percentile HHL, Maximum probability HLH, Robust mean absolute deviation HHL, difference entropy LHH, normalized grayscale unevenness LHH for grayscale size region matrix, maximum probability LHH, region percentage HHL, uniformity LHH, interquartile range HHL, 90th percentile LHH, dependency variance HHL, robust mean absolute deviation LHH, interquartile range LHH, run entropy LHH, small dependency high grayscale advantage HHL, sum entropy HLH, contrast LHH for grayscale co-occurrence matrix, 10th percentile LHH, grayscale variance HHL for grayscale run length matrix, clustering tendency HLH, run entropy HHH, small area advantage HLH, difference entropy HLL, normalized grayscale unevenness HHL for grayscale dependency matrix, large dependency low grayscale advantage HHL,Normalized size region unevenness HLH, inverse difference HLH, long run advantage HLH, sum of squares HLH, run entropy HHL, small area advantage HLL, inverse difference moment HLH, normalized size region unevenness HLL, short run advantage HLH, small dependence advantage HLL, difference variance HLL, large dependence advantage HLL, difference mean LHH, difference variance HLH, grayscale variance HLH for grayscale dependency matrix, root mean square HLH, run percentage HLL, short run advantage HLL, variance HLH, grayscale variance HLH for grayscale size region matrix, long run advantage HLL, normalized run length unevenness HLL, region percentage HLL, 10th percentile HLH, cluster prominence HLH H, Normalized Dependency Non-Uniformity HLL, Entropy HLH, Normalized Grayscale Non-Uniformity LHH for Grayscale Dependency Matrix, Small Dependency Advantage HLH, Difference Mean HLH, Mean Absolute Deviation HLH, Run Variance HLH, Run Variance HLL, Robust Mean Absolute Deviation HLH, Normalized Grayscale Non-Uniformity HLH for Grayscale Size Area Matrix, Uniformity HLH, Interquartile Range HLH, Joint Entropy HLL, Inverse Moment LHH, Joint Energy HLL, Large Area Advantage HLL, Small Dependency Advantage, Complexity HHL, Dependency Variance HLL, Large Area Advantage HHL, 90th Percentile HLH, Inverse Difference LHH, Run Percent HLH, Run Variance, Area Percent, for Grayscale Co-occurrence Matrix Contrast HLH, long run advantage, large area advantage, normalized run length unevenness HLH, normalized size area unevenness, small area advantage, large dependence advantage HLH, normalized dependence unevenness, large dependence advantage, run percentage, short run advantage, area percentage HLH, area variance HLL, contrast for gray level co-occurrence matrix HLL, large area low gray level advantage HLH, area variance, difference mean HLL, gray level variance LHH for gray level run length matrix, inverse difference HLL, dependency entropy HHH, difference entropy, inverse difference moment HLL, joint energy LHL, joint energy LLH, normalized run length unevenness, area variance HHL, difference entropy LHL, small area high gray level advantage HH L, maximum probability HLL, grayscale variance HLH for grayscale run length matrix, inverse variance, dependency entropy LHH, normalized grayscale unevenness HLH_GLSDM, inverse difference LHL, inverse difference moment LHL, small dependence high grayscale advantage HLH, normalized dependency unevenness HLH, joint entropy LHL, long run advantage LHL, run variance LHL, inverse difference moment LLH, joint entropy LLH, large dependency advantage LHL, dependency variance, dependency variance LHL, inverse difference LLH, maximum probability LHL, difference mean, run entropy HLH, normalized dependency unevenness LHL, difference entropy LLH, large area advantage LHL, maximum probability LLH, contrast for grayscale co-occurrence matrix, run percentage LHL,Short run advantage LHL, short run high grayscale advantage HHL, sum entropy LLH, long run low grayscale advantage HHL, short run advantage LHH, small dependence advantage LHH, area percentage LHL, area variance LHL, inverse difference, inverse difference moment, small dependence advantage LHL, area percentage LHH, inverse variance HLL, large dependence low grayscale advantage HLH, difference mean LHL, small area high grayscale advantage HHH, run variance LLH, difference variance LHL, large area low grayscale advantage HLL, normalized run length unevenness LHL, contrast for gray level co-occurrence matrix LHL, dependency variance HLH, normalized inverse difference, maximum value LLL, normalized run length unevenness LHH, normalized inverse difference moment, long Run advantage LLH, size area unevenness HHH, normalized size area unevenness LHL, small area advantage LHL, interquartile range LHL, normalized grayscale unevenness LHL for grayscale size area matrix, run percentage LHH, small area high grayscale advantage HLH, uniformity LHL, difference mean LLH, difference variance, large area high grayscale advantage, long run advantage LHH, small dependence high grayscale advantage HHH, high grayscale run advantage HHL, range HHL, robust mean absolute deviation LHL, high grayscale advantage HHL, inverse variance LHL, inverse variance LLH, uniformity HLL, normalized grayscale unevenness HLL_GLSDM, large dependence advantage LLH, entropy LHL, sum of squares L LH, interquartile range HLL, interquartile range LLH, robust mean absolute deviation HLL, sum entropy LHL, 90th percentile LLH, complexity, normalized dependency non-uniformity LLH, normalized grayscale non-uniformity LHL for grayscale dependency matrix, robust mean absolute deviation LLH, run percentage LLH, small dependency advantage LLH, entropy HLL, high grayscale area advantage HHL, short run advantage LLH, uniformity LLH, area percentage LLH, dependency variance LLH, normalized grayscale non-uniformity LLH for grayscale size area matrix, mean absolute deviation LHL, sum of squares LHL, contrast LLH for grayscale co-occurrence matrix, large area low grayscale advantage LHH, entropy LLH, aggregation Class tendency LLH, mean absolute deviation LLH, normalized large and small area unevenness LLH, small area advantage LLH, complexity HLH, high grayscale run advantage HHH, large area advantage LLH, large dependence advantage LHH, difference variance LLH, correlation information measure 1, large area low grayscale advantage LHL, normalized run length unevenness LLH, run variance LHH, complexity HLL, large dependence advantage LLL, normalized dependence unevenness LLL, short run high grayscale advantage HLH, run entropy LHL, small dependence advantage LLL, dependence variance LLL, normalized grayscale unevenness HLL for grayscale dependency matrix, normalized grayscale unevenness LLH for grayscale dependency matrix,Grayscale variance LHL for grayscale dependency matrix, high grayscale area advantage HHH, large area advantage HLH, large area advantage LLL, long run advantage LLL, mean absolute deviation HLL, run percentage LLL, short run advantage LLL, area percentage LLL, area variance LLH, normalized run length unevenness LLL, variance LHL, complexity LHL, grayscale variance LHL for grayscale size area matrix, long run low grayscale advantage HLH, run square Difference LLL, regional variance LLL, range HHH, small area advantage LLL, sum of squares HLL, variance HLL, grayscale variance HLL for grayscale size region matrix, grayscale variance HLL for grayscale dependency matrix, inverse difference moment LLL, normalized size region non-uniformity HHH, normalized size region non-uniformity LLL, small dependency high grayscale advantage LHH, inverse difference LLL, minimum value HHH, minimum value HHL, run entropy LLH, small area advantage HHH, 10th percentile HLL, inverse variance LLL, high grayscale advantage HHH, root mean square LLH, small dependence high grayscale advantage HLL, grayscale variance LHL for grayscale run length matrix, grayscale variance LLH for grayscale dependency matrix, autocorrelation HHL, variance LLH, 10th percentile LHL, maximum value HHL, short run high grayscale advantage HHH, grayscale variance LLH for grayscale size region matrix, sum entropy HLL, difference entropy LLL, clustering Tendency LHL, regional variance HLH, difference mean LLL, large dependence low gray advantage LHL, root mean square HLL, complexity LHH, high gray advantage HLH, large dependence low gray advantage HLL, high gray run advantage HLH, root mean square LHL, small dependence high gray advantage LHL, dependence entropy HHL, run entropy HLL, large dependence low gray advantage LHH, large and small area unevenness HHL, clustering tendency HLL and maximum probability HHH; Group 13 consists of difference entropy LHH, contrast LHH for gray-level co-occurrence matrix, difference mean LHH, joint entropy LHH, difference variance LHH, sum of squares LHH, entropy LHH, mean absolute deviation LHH, gray-level variance LHH for gray-level dependence matrix, robust mean absolute deviation LHH, root mean square LHH, sum entropy LHH, variance LHH, gray-level variance LHH for gray-level size area matrix, interquartile range LHH, clustering tendency LHH, 90th percentile LHH, run entropy LHH, short run advantage LHH, normalized run length unevenness LHH, cluster prominence LHH, small dependence advantage LHH, run percentage LHH, interquartile range HHH, robust mean absolute deviation HH H, 90th percentile HHH, mean absolute deviation HHH, grayscale variance LHH for grayscale run length matrix, root mean square HHH, variance HHH, area percentage LHH, joint entropy HHH, dependence entropy LHH, difference entropy HHH, sum of squares HHH, entropy HHH, difference variance HHH, grayscale variance HHH for grayscale size area matrix HHH, grayscale variance HHH for grayscale dependency matrix, difference entropy LLH, difference entropy LHL, sum entropy HHH, grayscale variance HHH for grayscale run length matrix, cluster prominence HHH, difference mean LLH, run percentage LHL, difference mean LHL, short run advantage LHL, normalized dependency unevenness LHL, small dependency advantage Potential LHL, Area Percentage LHL, Joint Entropy LLH, Normalized Size Area Inhomogeneity LHH, Normalized Size Area Inhomogeneity LHL, Small Area Advantage LHL, Contrast LHL for Gray Level Co-occurrence Matrix, Normalized Run Length Inhomogeneity LHL, Contrast LLH for Gray Level Co-occurrence Matrix, Difference Variance LHL, Run Percentage LLH, Short Run Advantage LLH, Small Area Advantage LHH, Joint Entropy LHL, Normalized Dependence Inhomogeneity LLH, Small Area Advantage LLH, Small Dependence Advantage LLH, Normalized Size Area Inhomogeneity LLH, Difference Variance LLH, Area Percentage LLH, Small Dependence High Gray Level Advantage LHH, Interquartile Range LLH, Robust Mean Absolute Deviation LLH, Normalized run length unevenness LLH, complexity LHH, small dependency advantage, run percentage, sum of squares LLH, area percentage, entropy LLH, mean absolute deviation LLH, short run advantage, normalized dependency unevenness, sum entropy LLH, normalized run length unevenness, normalized size area unevenness, small area advantage, interquartile range LHL, robust mean absolute deviation LHL, difference entropy, short run advantage HHL, small area advantage HHL, normalized size area unevenness HHL, entropy LHL, size area unevenness HHH, difference mean, run entropy LLH, normalized run length unevenness HHL, run percentage HHL, clustering tendency HHH, 90th percentile LLH,Grayscale variance LLH for grayscale dependency matrix, small area advantage HHH, variance LLH, mean absolute deviation LHL, small dependence advantage HHL, contrast for grayscale co-occurrence matrix, small area advantage HLH, clustering tendency LLH, grayscale variance LLH for grayscale size region matrix, sum of squares LHL, normalized dependency unevenness HHL, root mean square LLH, normalized size region unevenness HHH, normalized size region unevenness HLH, complexity LHL, difference entropy HHL, correlation information measure 1, total energy LHH, run entropy LHL, region percentage HHL, difference mean H HL, Sum Entropy LHL, Grayscale Variance LHL for Grayscale Dependency Matrix, Variance LHL, Grayscale Variance LHL for Grayscale Size Area Matrix, Contrast HHL for Grayscale Co-occurrence Matrix, Difference Variance, Difference Variance HHL, Short Run Advantage HLH, Joint Entropy HHL, Small Area High Grayscale Advantage HHH, Small Dependence High Grayscale Advantage LHL, Difference Entropy HLH, Small Dependence Advantage HLH, Grayscale Variance LLH for Grayscale Run Length Matrix, Normalized Dependence Non-uniformity LLL, Grayscale Variance LHL for Grayscale Run Length Matrix, Interquartile Range HHL, Run Percentage LLL, Area Percentage L LL, robust mean absolute deviation HHL, short run advantage LLL, size area unevenness LHH, small dependence advantage LLL, normalized run length unevenness LLL, complexity, run percentage HLH, contrast for gray-level co-occurrence matrix HHH, maximum value LHH, mean absolute deviation HHL, entropy HHL, normalized size area unevenness LLL, small area advantage LLL, difference mean HLH, sum of squares HHL, 90th percentile HHL, root mean square HHL, normalized run length unevenness HLH, variance HHL, gray-level variance for gray-level size area matrix HHL, The gray level dependency matrix consists of gray level variance HHL, root mean square LHL, joint entropy HLH, run entropy HHH, area percentage HLH, total energy HHH, clustering tendency LHL, sum entropy HHL, energy LHH, contrast HLH for gray level co-occurrence matrix, difference variance HLH, normalized dependency unevenness LHH, small dependency high gray level advantage HHL, clustering tendency HHL, short run high gray level advantage LHH, normalized dependency unevenness HLH, robust mean absolute deviation HLH, small area high gray level advantage LHL, interquartile range HLH, range LHH and small dependency high gray level advantage HHH; Group 14 consists of the sum of squares LHH, cluster tendency LHH, entropy LHH, grayscale variance LHH for grayscale dependency matrix, joint entropy LHH, grayscale variance LHH for grayscale size region matrix, mean absolute deviation LHH, root mean square LHH, variance LHH, contrast LHH for grayscale co-occurrence matrix, difference entropy LHH, difference variance LHH, sum entropy LHH, difference mean LHH, robust mean absolute deviation LHH, interquartile range LHH, 90th percentile LHH, run entropy LHH, cluster prominence LHH, short run advantage LHH, small dependence advantage LHH, normalized run length unevenness LHH, run percentage LHH, grayscale variance LHH for grayscale run length matrix, Interquartile range HHH, robust mean absolute deviation HHH, 90th percentile HHH, dependence entropy LHH, mean absolute deviation HHH, area percentage LHH, root mean square HHH, variance HHH, joint entropy HHH, entropy HHH, difference entropy HHH, sum of squares HHH, grayscale variance for grayscale size area matrix HHH, grayscale variance for grayscale dependency matrix HHH, difference variance HHH, run percentage LHL, difference entropy LHL, normalized dependency unevenness LHL, short run advantage LHL, small dependency advantage LHL, area percentage LHL, difference entropy LLH, normalized size area unevenness LHL, small area advantage LHL, difference mean LHL, normalized run length Non-uniformity LHL, Difference Mean LLH, Sum Entropy HHH, Contrast LHL for gray-level co-occurrence matrix, Cluster Prominence HHH, Difference Variance LHL, Run Percentage LLH, Small Area Advantage LLH, Small Dependence Advantage, Short Run Advantage LLH, Normalized Size Area Non-uniformity LLH, Small Dependence Advantage LLH, Area Percentage, Contrast LLH for gray-level co-occurrence matrix, Normalized Dependence Non-uniformity LLH, Joint Entropy LLH, Run Percentage, Area Percentage LLH, Short Run Advantage, Normalized Dependence Non-uniformity, Gray Variance HHH for gray-level run length matrix, Joint Entropy LHL, Mean Absolute Deviation LLH, Robust Mean Absolute Deviation LLH, Normalized Size Area Non-uniformity degree, small area advantage, interquartile range LLH, entropy LLH, normalized run length unevenness LLH, normalized run length unevenness, sum of squares LLH, difference variance LLH, small dependence high gray level advantage LHH, interquartile range LHL, robust mean absolute deviation LHL, complexity LHH, difference entropy, entropy LHL, gray level variance LLH for gray level dependence matrix, sum entropy LLH, variance LLH, difference mean, short run advantage HHL, run entropy LLH, small area advantage HHL, gray level variance LLH for gray level size area matrix, normalized size area unevenness HHL, root mean square LLH, 90th percentile LLH, contrast for gray level co-occurrence matrix, mean absolute deviation LHL,Run percentage HHL, clustering tendency LLH, normalized run length unevenness HHL, small dependence advantage HHL, normalized size area unevenness LHH, sum of squares LHL, small area advantage HLH, normalized dependency unevenness HHL, clustering tendency HHH, run entropy LHL, normalized size area unevenness HLH, area percentage HHL, complexity LHL, grayscale variance LHL for grayscale size area matrix, variance LHL, grayscale variance LHL for grayscale dependency matrix, small area advantage LHH, sum entropy LHL, difference entropy HHL, difference variance, normalized dependency unevenness LLL, difference mean HHL, grayscale variance LLH for grayscale run length matrix, area percentage ratio LLL, run percentage LLL, size area unevenness HHH, small dependence advantage LLL, short run advantage LLL, normalized run length unevenness LLL, correlation information measure 1, normalized size area unevenness LLL, small area advantage LLL, complexity, contrast HHL for gray-level co-occurrence matrix, small dependence high gray-level advantage LHL, short run advantage HLH, small dependence advantage HLH, gray-level variance LHL for gray-level run length matrix, joint entropy HHL, difference variance HHL, interquartile range HHL, robust mean absolute deviation HHL, small area advantage HHH, mean absolute deviation HHL, root mean square LHL, run percentage HLH, normalized size area unevenness HHH, Entropy HHL, 90th percentile HHL, root mean square HHL, total energy LHH, variance HHL, difference entropy HLH, grayscale variance HHL for grayscale size region matrix, sum of squares HHL, grayscale variance HHL for grayscale dependency matrix, normalized run length unevenness HLH, region percentage HLH, clustering tendency LHL, difference mean HLH, small area high grayscale advantage HHH, maximum value LHH, difference mean LLL, sum entropy HHL, small dependence high grayscale advantage HHL, clustering tendency HHL, normalized dependency unevenness HLH, difference entropy LLL, joint entropy HLH, short run high grayscale advantage LHH, contrast HLH for grayscale co-occurrence matrix, run entropy HHH, large Small area unevenness LHH, small area high gray level advantage LHL, robust mean absolute deviation HLH, interquartile range HLH, difference variance HLH, range LHH, mean absolute deviation HLH, run entropy HHL, cluster prominence HHL, gray level variance HLH for gray level dependency matrix, entropy HLH, root mean square HLH, small dependence high gray level advantage HHH, variance HLH, gray level variance HHL for gray level run length matrix, gray level variance HLH for gray level size area matrix, energy LHH, small area high gray level advantage LHH, complexity HHL, contrast HHH for gray level co-occurrence matrix, 90th percentile HLH, sum of squares HLH, run percentage HLL, size area unevenness HHL,The complexity LLH, normalized dependency heterogeneity HLL, small dependency advantage HLL, 90th percentile LHL, normalized dependency heterogeneity LHH and area percentage HLL are composed; and, Group 15 consists of grayscale variance LHH, root mean square LHH, variance LHH, entropy LHH, mean absolute deviation LHH, sum of squares LHH for grayscale dependence matrix, normalized grayscale unevenness LHH, clustering tendency LHH, uniformity LHH for grayscale co-occurrence matrix, contrast LHH, difference variance LHH, run entropy LHH, 90th percentile LHH, joint entropy LHH, robust mean absolute deviation LHH, 10th percentile LHH, interquartile range LHH, sum entropy LHH, difference mean LHH, joint energy LHH, inverse moment LHH, inverse difference LHH, maximum probability LHH, cluster prominence LHH for grayscale size area matrix. Uniformity LHH, Grayscale variance LHH for grayscale size region matrix, Short run advantage LHH, Dependence entropy LHH, Small dependence advantage LHH, Normalized run length unevenness LHH, Long run advantage LHH, Interquartile range HHH, Mean absolute deviation HHH, 90th percentile HHH, Robust mean absolute deviation HHH, Root mean square HHH, Run percentage LHH, Variance HHH, 10th percentile HHH, Region percentage LHH, Normalized grayscale unevenness HHH for grayscale dependency matrix, Uniformity HHH, Entropy HHH, Run variance LHH, Large dependence advantage LHH, Grayscale variance HHH for grayscale size region matrix, Grayscale variance for grayscale dependency matrix HHH, joint entropy HHH, sum of squares HHH, difference entropy HHH, difference variance HHH, joint energy HHH, large area low grayscale advantage LHH, run variance LHL, long run advantage LHL, inverse difference moment LHL, inverse difference LHL, inverse difference moment LLH, inverse difference LLH, large dependence advantage LHL, run percentage LHL, small dependence advantage LHL, area percentage LHL, normalized dependence unevenness LHL, short run advantage LHL, normalized size area unevenness LHL, small area advantage LHL, difference entropy LHL, cluster prominence HHH, normalized grayscale unevenness for grayscale size area matrix HHH, difference mean LHL, run variance LLH, difference entropy LLH, large area Advantage LHL, Long run advantage LLH, Small dependency high gray level advantage LHH, Sum entropy HHH, Dependence variance LHL, Inverse variance LLH, Inverse variance LHL, Gray level variance HHH for gray level run length matrix, Normalized run length non-uniformity LHL, Contrast LHL for gray level co-occurrence matrix, Difference mean LLH, Region variance LHL, Complexity LHH, Run variance, Normalized size region non-uniformity LLH, Small area advantage LLH, Difference variance LHL, Contrast LLH for gray level co-occurrence matrix, Long run advantage, Small dependency advantage LLH, Mean absolute deviation LLH, Run percentage LLH, Small dependency advantage, Joint energy LLH, Short run advantage LLH, Region percentage,Normalized dependency non-uniformity LLH, area percentage LLH, entropy LLH, joint energy LHL, large dependency advantage LLH, robust mean absolute deviation LLH, interquartile range LLH, uniformity LLH, normalized grayscale non-uniformity LLH for grayscale dependency matrix, run percentage, small area advantage, difference variance LLH, joint entropy LLH, large area advantage, normalized size area non-uniformity, sum of squares LLH, short run advantage, normalized grayscale non-uniformity LLH for grayscale size area matrix, normalized dependency non-uniformity, joint entropy LHL, large dependency advantage, normalized run length non-uniformity LLH, large area low grayscale advantage LHL, large dependency low grayscale advantage LHH, maximum probability LHL , Uniformity LHL, Normalized Grayscale Nonuniformity LHL for Grayscale Dependency Matrix, Grayscale Variance LLH for Grayscale Dependency Matrix, Normalized Run Length Nonuniformity, Variance LLH, Area Variance, Interquartile Range LHL, Long Run Advantage HHL, Robust Mean Absolute Deviation LHL, Grayscale Variance LLH for Grayscale Size Area Matrix, Large Area Advantage LLH, Small Area Advantage HHL, Dependence Variance LLH, Maximum Probability LLH, Run Entropy LLH, Normalized Size Area Nonuniformity HHL, Entropy LHL, Root Mean Square LLH, Normalized Size Area Nonuniformity LHH, Run Variance HHL, Difference Mean, Normalized Grayscale Nonuniformity LHL for Grayscale Size Area Matrix, Inverse Moment, Inverse Variance, Short run advantage HHL, inverse difference, difference entropy, mean absolute deviation LHL, 90th percentile LLH, clustering tendency LLH, contrast for gray-level co-occurrence matrix, small area advantage HLH, sum entropy LLH, complexity LHL, small dependence advantage HHL, regional variance LLH, normalized inverse difference, large dependence advantage HHL, run percentage HHL, normalized size regional unevenness HLH, large area advantage LHH, normalized run length unevenness HHL, 10th percentile LHL, sum of squares LHL, dependency variance, normalized inverse difference moment, small area advantage LHH, clustering tendency HHH, run entropy LHL, variance LHL, gray-level variance LHL for gray-level size regional matrix, regional percentage Ratio HHL, grayscale variance LLH for grayscale run length matrix, normalized dependency non-uniformity HHL, large dependency low grayscale advantage LHL, inverse difference HHL, inverse difference moment HHL, large area high grayscale advantage, small dependency high grayscale advantage LHL, large area low grayscale advantage HHL, large and small area non-uniformity HHH, dependency variance HHL, area variance LHH, difference entropy HHL, grayscale variance LHL for grayscale dependency matrix, difference mean HHL, long run advantage HLH, difference variance, 10th percentile LLH, long run advantage LLL, sum entropy LHL, complexity, maximum probability HHL, run variance LLL, normalized dependency non-uniformity LLL, area percentage LLL, inverse difference HLH,Run percentage LLL, small dependence advantage HLH, small dependence advantage LLL, large dependence advantage LLL, short run advantage HLH, short run advantage LLL, small area advantage HHH, contrast for gray-level co-occurrence matrix HHL, large area advantage LLL, joint energy HHL, maximum value LHH, normalized run length unevenness LLL, run variance HLH, normalized size area unevenness HHH, 10th percentile HHL, inverse difference moment HLH, normalized size area unevenness LLL, small area advantage LLL, small area high gray level advantage HHH, interquartile range HHL, difference variance HHL, robust mean absolute deviation HHL, root mean square LHL, short run high gray level advantage LHH, inverse difference moment L LL, Large Area Dominance HHL, Dependence Variance LLL, Joint Entropy HHL, Uniformity HHL, Normalized Grayscale Nonuniformity HHL for Grayscale Dependence Matrix, Regional Variance LLL, Inverse Variance LLL, Mean Absolute Deviation HHL, Entropy HHL, Correlation Information Measure 1, Total Energy LHH, Inverse Variance LLL, Range LHH, Root Mean Square HHL, Run Percentile HLH, Variance HHL, 90th Percentile HHL, Grayscale Variance HHL for Grayscale Dependence Matrix, Difference Mean HLH, Grayscale Variance HHL for Grayscale Size Region Matrix, Difference Entropy HLH, Large Dependence Dominance HLH, Region Percentage HLH, Normalized Run Length Nonuniformity HLH, Region Variance HHL, Run Entropy HH H, small area high gray level advantage LHL, square sum HHL, clustering tendency LHL, long run low gray level advantage LHH, maximum probability HLH, small area high gray level advantage LHH, small dependence high gray level advantage HHL, joint energy HLH, contrast HLH for gray level co-occurrence matrix, difference mean LLL, small dependence high gray level advantage HHH, large dependence low gray level advantage HHL, sum entropy HHL, 10th percentile HLH, clustering tendency HHL, normalized dependency unevenness HLH, normalized gray level unevenness HHL for gray level size area matrix, high gray level advantage LHH, high gray level run advantage LHH, joint entropy HLH, robust mean absolute deviation HLH, interquartile range HLH, size area Non-uniformity LHH, Difference Entropy LLL, Difference Variance HLH, Run Entropy HHL, Contrast HHH for Gray-level Co-occurrence Matrix, Gray-level Variance HHL for Gray-level Run Length Matrix, Gray-level Variance HLH for Gray-level Size Area Matrix, Mean Absolute Deviation HLH, Root Mean Square HLH, Uniformity HLH, Variance HLH, Entropy HLH, Normalized Gray-level Non-uniformity HLH for Gray-level Dependence Matrix, Gray-level Variance HLH for Gray-level Dependence Matrix, Complexity HHL, Large Area Low Gray-level Dominance HLH, Large Dependence High Gray-level Dominance, Cluster Prominence HHL, Complexity LLH, 90th Percentile HLH, Energy LHH, Short Run High Gray-level Dominance LHL, Sum of Squares HLH, Dependence Variance HLH,It consists of short run high grayscale advantage HHH, high grayscale advantage LHL, high grayscale run advantage LHL, correlation, run variance HLL, high grayscale area advantage LHH, long run advantage HLL and small dependence high grayscale advantage HLH.

9. The method of claim 8, wherein the at least two epicardial radiomics features comprise at least two of the following: inverse difference moment HHH, minimum value LHH, low grayscale area advantage LLL, maximum probability LLL, busyness LHH, regional entropy LLL, run entropy LLL, maximum 3D diameter, long run high grayscale advantage LLL, joint average LLL, difference entropy LLL, sum entropy HHH, difference entropy LHH, sum of squares LHH, and grayscale variance LHH for grayscale dependency matrix.

10. The method of claim 8, wherein the at least two epicardial radiomics features are selected from epicardial radiomics features of clusters A to D, wherein the at least two epicardial radiomics features are each selected from a different cluster, and wherein: Cluster A consisted of epicardial radiomic features from groups 1 to 5; Cluster B consisted of epicardial radiomic features of groups 6 and 7; Cluster C consisted of epicardial radiomic features of group 8; and Cluster D consisted of epicardial radiomic features from clusters 9 to 15.

11. The method according to claim 1 or 2, wherein: The plurality of epicardial radiomics features include at least two epicardial radiomics features selected from epicardial radiomics features of clusters A to D, wherein the at least two epicardial radiomics features are each selected from a different cluster, and wherein: Cluster A is composed of inverse moment HHH, minimum value LHH, regional variance LLL, low grayscale regional advantage LLL, maximum probability LLL, elongation, cluster shadow LLL, busyness LHH, grayscale unevenness LLL and skewness HHH; Cluster B consists of regional entropy LLL, cluster prominence LLL, grayscale variance LLL for grayscale dependency matrix, and run entropy LLL; Cluster C consists of the smallest axis, the largest 2D diameter row, the major axis, the largest 2D diameter column, the largest 2D diameter slice, and the largest 3D diameter; and Cluster D consists of autocorrelation LLH, long run high grayscale advantage LLL, joint mean LLL, autocorrelation LLL, difference entropy LLL, difference entropy LLH, sum of squares HLH, sum of squares HHH, sum entropy HHH, grayscale variance LLH for grayscale dependency matrix, difference entropy LHH, sum of squares LHH, and grayscale variance LHH for grayscale dependency matrix.

12. The method according to claim 11, wherein: Cluster A is composed of inverse moment HHH, minimum value LHH, low gray area advantage LLL, maximum probability LLL and busyness LHH; Cluster B is composed of region entropy LLL and run entropy LLL; Cluster C consists of the largest 3D diameters; and Cluster D consists of the long-run high grayscale dominance LLL, the joint mean LLL, the difference entropy LLL, the sum entropy HHH, the difference entropy LHH, the sum of squares LHH and the grayscale variance LHH for the grayscale dependency matrix.

13. The method of claim 1 or 2, further comprising identifying the epicardial region using an automatic segmentation algorithm.

14. The method according to claim 13, wherein: Machine learning has been used to train the segmentation algorithm to segment the medical imaging data.

15. The method according to claim 1 or 2, wherein: The epicardial region is composed of voxels of the medical imaging data, the attenuation values ​​of the medical imaging data falling within a given range of attenuation values.

16. The method according to claim 15, wherein: The given range of attenuation values ​​corresponds to fat.

17. The method according to claim 15, wherein: The given range of attenuation values ​​corresponds to connective tissue.

18. The method according to claim 15, wherein: The given range includes attenuation values ​​of -190 Hounsfield units to -30 Hounsfield units.

19. The method according to claim 15, wherein: The given range includes attenuation values ​​of -190 Hounsfield units to +150 Hounsfield units.

20. The method of claim 1 or 2, further comprising predicting a subject's risk of developing a cardiac condition or experiencing a stroke based at least on the calculated value of the epicardial radiomic signature.

21. The method according to claim 20, wherein: The cardiac condition is arrhythmia.

22. The method of claim 1 or 2, further comprising identifying whether an individual is at risk for stroke based on the calculated value of the epicardial radiomic signature.

23. The method of claim 22, further comprising: If the individual is identified as being at risk for stroke, a preventive treatment is administered or prescribed to the individual to reduce the risk of stroke.

24. The method of claim 1 or 2, further comprising determining whether the subject has myocardial disease based at least on the calculated value of the epicardial radiomic signature.

25. The method according to claim 24, wherein: The myocardial disease is fibrosis.

26. The method according to claim 24, wherein: The myocardial disease is inflammation.

27. The method according to claim 24, wherein: The myocardial disease is oxidative stress.

28. A method for deriving an epicardial radiomic signature indicative of cardiac health, the method comprising constructing an epicardial radiomic signature indicative of cardiac health using a radiomics dataset, the epicardial radiomic signature calculated based on a second plurality of epicardial radiomic features; in, The data set includes values ​​of a first plurality of epicardial radiomic features obtained from medical imaging data of an epicardial region including epicardial tissue for each individual of a plurality of individuals, the plurality of individuals including a first group of individuals having a cardiac condition or myocardial disease or having a history of stroke and a second group of individuals not having a cardiac condition or myocardial disease or having a history of stroke; wherein the second plurality of epicardial radiomic features are selected from the first plurality of epicardial radiomic features; and Wherein the epicardial radiomic signature is configured to provide a measure of texture of the epicardial tissue.

29. The method according to claim 28, wherein: The method further includes identifying significant epicardial radiomic features from the first plurality of epicardial radiomic features that are each significantly associated with the cardiac condition or the myocardial disease or the history of stroke, the second plurality of epicardial radiomic features including at least two epicardial radiomic features that are different significant epicardial radiomic features or are co-linear therewith.

30. The method of claim 29, wherein: The method further comprises identifying a subset of the significant epicardial radiomic features using a feature selection machine learning algorithm, wherein the at least two epicardial radiomic features are different significant epicardial radiomic features belonging to the subset or are co-linear therewith.

31. The method according to claim 30, wherein: The at least two epicardial radiomic features include all significant epicardial radiomic features belonging to the subset or their co-linear equivalents.

32. The method according to claim 29 or 30, further comprising: Groups of epicardial radiomic features are identified, each of the groups comprising the significant epicardial radiomic feature and one of its collinear equivalents that is collinear with the significant epicardial radiomic feature, the at least two epicardial radiomic features being selected from different groups.

33. The method according to claim 29 or 30, wherein: The method includes identifying a plurality of clusters of the significant epicardial radiomic features by performing a cluster analysis, and wherein each of the at least two epicardial radiomic features is selected from a different cluster or is selected to be co-linear with a significant epicardial radiomic feature from a different cluster.

34. The method of claim 33, wherein: The cluster analysis identifies the clusters based on the strength of correlations between the significant epicardial radiomic features.

35. A method according to any one of claims 28 to 31, wherein: The epicardial radiomic signature is configured to be correlated with the cardiac condition or the myocardial disease or the history of stroke.

36. The method of claim 35, wherein: The epicardial radiomic signature is constructed to be significantly associated with the cardiac condition or the myocardial disease or the history of stroke.

37. A method according to any one of claims 28 to 31, wherein: The step of constructing the epicardial radiomic signature is performed using a machine learning algorithm.

38. A method according to any one of claims 28 to 31, wherein: The cardiac condition is associated with heart muscle health or disease.

39. The method of claim 38, wherein: The cardiac condition is arrhythmia.

40. The method of any one of claims 28 to 31, further comprising a system configured to calculate a value of the epicardial radiomics signature of a patient.

41. The method of any one of claims 28 to 31, further comprising characterizing an epicardial region of the patient by calculating a value of the epicardial radiomics signature for the epicardial region of the patient.

42. A method according to any one of claims 28 to 31, wherein: The epicardial region includes epicardial adipose tissue.

43. A method according to any one of claims 28 to 31, wherein: The epicardial region includes connective tissue.

44. A method according to any one of claims 28 to 31, wherein: The epicardial radiomics signature comprises a decision tree.

45. The method of claim 44, wherein: The epicardial radiomics signature includes a regression tree.

46. ​​A system configured to perform the method according to any one of claims 1 to 45.

Images