Radiomic signature of perivascular regions

Abstract

A method for characterizing a perivascular region using medical imaging data of a subject. The method includes calculating a value of a radiomic signature of the perivascular region using the medical imaging data. A method for deriving a radiomic signature for predicting cardiovascular risk is also disclosed. The method includes obtaining a radiomics dataset and constructing a perivascular radiomic signature using the 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 perivascular areas, 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] Despite significant progress in both primary and secondary cardiovascular prevention, coronary artery disease (CAD) remains a major cause of morbidity and mortality. Noninvasive diagnostic tests that assess the presence of CAD in patients presenting typical or atypical symptoms, such as coronary computed tomography angiography (CCTA), are the mainstay of modern cardiovascular diagnostics, especially in individuals with a low to intermediate pretest likelihood of CAD. Traditionally, such techniques rely on the detection of obstructive lesions or the presence and extent of coronary artery calcification for cardiovascular risk stratification. However, despite optimal medical therapy, a large number of patients still have elevated residual cardiovascular risk. Residual vascular inflammation, in particular, is a driver of adverse events, leading to both the formation and instability of atherosclerotic plaques, but can be difficult to diagnose using conventional tests, such as circulating inflammatory biomarkers that are nonspecific for vascular disease.

[0003] Traditionally, interpretation of clinical imaging studies has relied on subjective, operator-dependent, qualitative assessments of the imaged anatomy. While this approach is useful in a busy clinical environment, it ignores the vast amount of information contained in all clinical scans.

[0004] WO 2016 / 024128 A1 and WO 2018 / 078395 A1 define specific CCTA-based metrics, namely, fat attenuation index (FAI) or perivascular FAI (FAI PVAT ), a metric that reflects the normalized weighted average attenuation of the perivascular region surrounding human coronary arteries, was found to be a sensitive and dynamic biomarker of coronary inflammation and subsequently identified as a strong and independent predictor of adverse cardiac events. In the presence of vascular inflammation, the release of proinflammatory molecules from the diseased vessel wall inhibits differentiation and lipid accumulation in preadipocytes within the perivascular tissue (PVT), resulting in smaller, less differentiated, and lipid-free adipocytes. This is associated with a shift in the radiodensity of the PTV (measured as attenuation values) on computed tomography (CT) imaging from larger negative (approaching -190) to smaller negative (approaching -30) Hounsfield unit (HU) values, which may be captured by the FAI. Although both the biological significance and clinical value of the FAI have been extensively validated, it is insufficient to describe the full range of phenotypic variability observed in coronary PVT.

[0005] In a recent study, Kolossvary et al. showed that radiomic features can reliably identify plaques with the napkin-ring sign, a characteristic of high-risk plaques, which typically rely on qualitative assessment of plaque anatomy and the operator's experienced eye (Kolossvary M, Karady J, Szilveszter B, et al., Radiomic Features Are Superior to Conventional Quantitative Computed Tomographic Metrics to Identify Coronary Plaques With Napkin-Ring Sign, Circ Cardiovasc Imaging 2017;10(12):e006843). However, this approach focuses on phenotyping the plaque itself and ignores PVT and the valuable information that can be obtained from it.

[0006] What is needed is a method or tool that provides prognostic value for cardiovascular risk beyond current CCTA-based risk stratification tools, such as the presence and extent of CAD, the presence of high-risk plaque features, coronary artery calcium, and the recently described FAI. PVAT . Summary of the invention

[0007] According to a first aspect of the invention, there is provided a method for characterizing a perivascular region (e.g., its phenotype, e.g., composition and / or texture) using medical imaging data of a subject. The method may include calculating a value of a radiomics signature of the perivascular region using the medical imaging data. The radiomics signature may be calculated based on measured values ​​of at least two radiomics features of the perivascular region. The measured values ​​of the at least two radiomics features may be calculated based on or by the medical imaging data.

[0008] The radiomic signature may provide a measure of the texture of the perivascular region.

[0009] At least one of the at least two radiomic features may provide a measure of the texture of the perivascular region.

[0010] The radiomics signature can predict cardiovascular risk.

[0011] The radiomic signature can predict the likelihood that the subject will experience a major adverse cardiovascular event.

[0012] The radiomic signature can predict the likelihood of the subject experiencing a cardiac-specific major adverse cardiovascular event.

[0013] The radiomic signature may be indicative of cardiovascular health.

[0014] The radiomic signature may be indicative of vascular disease. For example, the radiomic signature may be indicative of vascular inflammation.

[0015] At least one of the at least two radiomic features may be calculated based on a wavelet transform of attenuation values.

[0016] Each of the at least two radiomic features may be volume independent and / or orientation independent.

[0017] The at least two radiomic features may be selected from the radiomic features of clusters 1 to 9, wherein each of the at least two radiomic features is selected from a different cluster, wherein:

[0018] Cluster 1 consists of short-run high grayscale dominance, high grayscale dominance, high grayscale run dominance, autocorrelation, mean sum, joint mean, and high grayscale region dominance;

[0019] Cluster 2 consists of skewness, skewness LLL, kurtosis, 90th percentile, 90th percentile LLL, median LLL, kurtosis LLL, and median;

[0020] Cluster 3 consists of run entropy, dependency entropy LLL, dependency entropy, regional entropy LLL, run entropy LLL and mean LLL;

[0021] Cluster 4 consists of small area low grayscale advantage, low grayscale area advantage, short run low grayscale advantage, low grayscale run advantage, low grayscale advantage, small dependence low grayscale advantage, grayscale variance LLL (GLSZM), grayscale variance (GLDM), variance, grayscale variance (GLDM), difference variance LLL, grayscale variance LLL (GLRLM), variance LLL, grayscale variance LLL (GLDM), sum of squares, contrast LLL, mean absolute deviation, interquartile range, robust mean absolute deviation, long run low grayscale advantage, difference variance, grayscale variance (GLSZM), normalized inverse moment, mean absolute deviation LLL, sum of squares LLL and contrast;

[0022] Cluster 5 consists of regional entropy, normalized grayscale nonuniformity (GLRLM), normalized grayscale nonuniformity LLL (GLRLM), sum entropy, joint energy, entropy, normalized grayscale nonuniformity (GLDM), joint energy, normalized grayscale nonuniformity LLL (GLDM), uniformity LLL, sum entropy LLL and uniformity;

[0023] Cluster 6 consists of regional entropy HHH, normalized size regional heterogeneity HHH, and small area advantage HHH;

[0024] Cluster 7 consists of intensity, roughness HLL, roughness, roughness LHL, roughness LLL, roughness LLH, roughness HHH, roughness HLH, roughness HHL, and roughness LHH;

[0025] Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, gray variance LLL (GLDM), variance LLL, gray variance LLL (GLRLM), gray variance (GLRLM), robust mean absolute deviation LLL, gray variance (GLDM), variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, gray variance LLL (GLSZM), sum of squares, robust mean absolute deviation, sum entropy, interquartile range, cluster prominence LLL, entropy LLL, 10th percentile LLL, 10th percentile; and

[0026] Cluster 9 consists of size area unevenness LLL, dependence unevenness HLL, grayscale unevenness HLL (GLSZM), grayscale unevenness (GLSZM), run length unevenness HHL, run length unevenness LHL, dependence unevenness LHL, dependence unevenness, run length unevenness HLH, busyness, run length unevenness LLH, dependence unevenness LLH, dependence unevenness LLL, size area unevenness, energy HLL, run length unevenness LHH, size area unevenness HLL, grayscale unevenness LLH (GLSZM), grayscale unevenness LHL (GLSZM), grayscale unevenness LLL (GLSZM), run length unevenness HLL, grayscale unevenness HLH (GLSZM), grayscale unevenness HHL (GLSZM), run length unevenness and run length unevenness HHH.

[0027] Clusters 1 to 9 can alternatively be defined as follows:

[0028] Cluster 1 consists of short-run high gray-level dominance, high gray-level dominance, high gray-level run dominance, autocorrelation, mean sum, and joint mean;

[0029] Cluster 2 consists of skewness, skewness LLL, kurtosis, 90th percentile, 90th percentile LLL, median LLL, and kurtosis LLL;

[0030] Cluster 3 consists of run entropy, dependency entropy LLL, dependency entropy, and regional entropy LLL;

[0031] Cluster 4 consists of small area low grayscale advantage, low grayscale area advantage, short run low grayscale advantage, low grayscale run advantage, low grayscale advantage, small dependence low grayscale advantage, grayscale variance LLL (GLSZM), grayscale variance (GLDM), variance, grayscale variance (GLDM), and difference variance LLL;

[0032] Cluster 5 consists of regional entropy, normalized gray level non-uniformity (GLRLM), and normalized gray level non-uniformity LLL (GLRLM);

[0033] Cluster 6 consists of the regional entropy HHH and the normalized size regional heterogeneity HHH;

[0034] Cluster 7 consists of intensity, roughness HLL, roughness, roughness LHL, roughness LLL, roughness LLH, and roughness HHH;

[0035] Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, grayscale variance LLL (GLDM), variance LLL, grayscale variance LLL (GLRLM), grayscale variance (GLRLM), robust mean absolute deviation LLL, grayscale variance (GLDM), variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, grayscale variance LLL (GLSZM), sum of squares, robust mean absolute deviation, sum entropy, interquartile range, cluster prominence LLL, entropy LLL, 10th percentile LLL and 10th percentile; and

[0036] Cluster 9 consists of size area unevenness LLL, dependence unevenness HLL, grayscale unevenness HLL (GLSZM), grayscale unevenness (GLSZM), run length unevenness HHL, run length unevenness LHL, dependence unevenness LHL, dependence unevenness, run length unevenness HLH, busyness, run length unevenness LLH, dependence unevenness LLH, dependence unevenness LLL, size area unevenness, energy HLL, run length unevenness LHH, size area unevenness HLL, grayscale unevenness LLH (GLSZM) and grayscale unevenness LHL (GLSZM).

[0037] Clusters 1 to 9 can alternatively be defined as follows:

[0038] Cluster 1 consists of short-run high gray-level dominance, high gray-level dominance, high gray-level run dominance, autocorrelation, mean sum, and joint mean;

[0039] Cluster 2 consists of skewness, skewness LLL, kurtosis, and 90th percentile;

[0040] Cluster 3 consists of run entropy, dependency entropy LLL, and dependency entropy;

[0041] Cluster 4 consists of small area low grayscale advantage, low grayscale area advantage, short run low grayscale advantage, low grayscale run advantage, low grayscale advantage, and small dependency low grayscale advantage;

[0042] Cluster 5 consists of regional entropy and normalized gray-level non-uniformity (GLRLM);

[0043] Cluster 6 consists of regional entropy HHH;

[0044] Cluster 7 consists of intensity, roughness HLL, roughness, roughness LHL, roughness LLL, and roughness LLH;

[0045] Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, grayscale variance LLL (GLDM), variance LLL, grayscale variance LLL (GLRLM), grayscale variance (GLRLM), robust mean absolute deviation LLL, grayscale variance (GLDM), variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, grayscale variance LLL (GLSZM), sum of squares, robust mean absolute deviation, sum entropy, interquartile range, cluster prominence LLL, entropy LLL, 10th percentile LLL and 10th percentile; and

[0046] Cluster 9 consists of size area unevenness LLL, dependence unevenness HLL, grayscale unevenness HLL (GLSZM), grayscale unevenness (GLSZM), run length unevenness HHL, run length unevenness LHL, dependence unevenness LHL, dependence unevenness, run length unevenness HLH, busyness, run length unevenness LLH, dependence unevenness LLH, dependence unevenness LLL and size area unevenness.

[0047] Clusters 1 to 9 can alternatively be defined as follows:

[0048] Cluster 1 consists of short-run high grayscale dominance, high grayscale dominance, high grayscale run dominance, and autocorrelation;

[0049] Cluster 2 consists of skewness and skewness LLL;

[0050] Cluster 3 consists of run entropy and dependency entropy LLL;

[0051] Cluster 4 consists of small areas with low grayscale dominance and low grayscale area dominance;

[0052] Cluster 5 consists of regional entropy;

[0053] Cluster 6 consists of regional entropy HHH;

[0054] Cluster 7 consists of intensity;

[0055] Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, grayscale variance LLL (GLDM), variance LLL, grayscale variance LLL (GLRLM), grayscale variance (GLRLM), robust mean absolute deviation LLL, grayscale variance (GLDM), variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, grayscale variance LLL (GLSZM), sum of squares, robust mean absolute deviation, sum entropy, interquartile range, cluster prominence LLL, entropy LLL, 10th percentile LLL and 10th percentile; and

[0056] Cluster 9 consists of large and small area non-uniformity LLL, dependence non-uniformity HLL, grayscale non-uniformity HLL (GLSZM), grayscale non-uniformity (GLSZM), run length non-uniformity HHL and run length non-uniformity LHL.

[0057] Clusters 1 to 9 can alternatively be defined as follows:

[0058] Cluster 1 consists of short runs with high grayscale dominance, high grayscale dominance, and high grayscale runs with dominance;

[0059] Cluster 2 consists of skewness and skewness LLL;

[0060] Cluster 3 consists of run entropy;

[0061] Cluster 4 consists of small areas with low grayscale dominance and low grayscale area dominance;

[0062] Cluster 5 consists of regional entropy;

[0063] Cluster 6 consists of regional entropy HHH;

[0064] Cluster 7 consists of intensity;

[0065] Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, gray variance LLL (GLDM), variance LLL, gray variance LLL (GLRLM), gray variance (GLRLM), robust mean absolute deviation LLL, gray variance (GLDM), variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, gray variance LLL (GLSZM), sum of squares, robust mean absolute deviation, and sum entropy; and

[0066] Cluster 9 consists of large and small regions of heterogeneity LLL.

[0067] Clusters 1 to 9 can alternatively be defined as follows:

[0068] Cluster 1 consists of short runs with high grayscale dominance, high grayscale dominance, and high grayscale runs with dominance;

[0069] Cluster 2 consists of skewness, skewness LLL, kurtosis, 90th percentile, median LLL, kurtosis LLL, and median;

[0070] Cluster 3 consists of run entropy, run entropy LLL, and median LLL;

[0071] Cluster 4 consists of small area low gray level dominance, low gray level region dominance, and gray level variance (GLSZM);

[0072] Cluster 5 consists of regional entropy, normalized gray level non-uniformity (GLRLM), and uniformity;

[0073] Cluster 6 consists of regional entropy HHH;

[0074] Cluster 7 consists of intensity;

[0075] Cluster 8 consists of cluster tendency LLL, cluster tendency, mean absolute deviation LLL, grayscale variance LLL (GLDM), variance LLL, grayscale variance LLL (GLRLM), robust mean absolute deviation LLL, grayscale variance LLL (GLSZM), interquartile range LLL, sum entropy LLL, grayscale variance (GLRLM), variance, grayscale variance (GLDM), mean absolute deviation, sum entropy, robust mean absolute deviation, interquartile range, entropy LLL, 10th percentile LLL, 10th percentile, entropy, uniformity LLL, normalized grayscale unevenness LLL (GLDM), normalized grayscale unevenness LLL (GLRLM), root mean square, normalized grayscale unevenness (GLDM), root mean square LLL and long run low grayscale advantage; and

[0076] Cluster 9 consists of LLL, busyness, and LLL size-area unevenness.

[0077] The at least two radiomic features may be selected from the following: median LLL, mean LLL, median, root mean square LLL, mean, kurtosis, root mean square, runs entropy LLL (GLRLM), uniformity, 90th percentile, normalized grayscale non-uniformity (GLRLM), uniformity LLL, skewness, normalized grayscale non-uniformity LLL (GLRLM), 10th percentile LLL, skewness LLL, 10th percentile, entropy, interquartile range LLL, robust mean absolute deviation LLL, runs entropy (GLRLM), interquartile range, sum entropy (GLCM), normalized grayscale non-uniformity LLL (GLRLM), dependent non-uniformity Kurtosis LHL (GLDM), kurtosis LLL, run length unevenness HHL (GLRLM), entropy LLL, robust mean absolute deviation, sum entropy LLL (GLCM), 90th percentile LLL, run entropy HHL (GLRLM), energy, energy LLL, intensity (NGTDM), autocorrelation (GLCM), mean absolute deviation LLL, high gray dominance (GLDM), joint mean (GLCM), sum of means (GLCM), short run high gray dominance (GLRLM), energy HHH, high gray run dominance (GLRLM), run entropy HHH (GLRLM), energy HHL, and mean absolute deviation.

[0078] The at least two radiomic features may be selected from the radiomic features of clusters 1 to 8. The at least two radiomic features may be selected from the radiomic features of clusters 1 to 7. The at least two radiomic features may be selected from the radiomic features of clusters 1 to 6. The at least two radiomic features may be selected from the radiomic features of clusters 1 to 5. The at least two radiomic features may be selected from the radiomic features of clusters 1 to 4. The at least two radiomic features may be selected from the radiomic features of clusters 1 to 3. The at least two radiomic features may be selected from the radiomic features of clusters 1 and 2.

[0079] The at least two radiomic features may include at least three radiomic features. The at least two radiomic features may include at least four radiomic features. The at least two radiomic features may include at least five radiomic features. The at least two radiomic features may include at least six radiomic features. The at least two radiomic features may include at least seven radiomic features. The at least two radiomic features may include at least eight radiomic features. The at least two radiomic features may include at least nine radiomic features.

[0080] The at least two radiomic features may include at least six radiomic features, wherein the six radiomic features are short run high grayscale dominance, skewness, run entropy, small area low grayscale dominance, regional entropy HHH, and regional entropy.

[0081] The at least two radiomic features may include at least nine radiomic features, wherein the nine radiomic features are short run high grayscale advantage, skewness, run entropy, small area low grayscale advantage, regional entropy HHH, regional entropy, intensity, clustering tendency LLL and large and small area unevenness.

[0082] The radiomic signature may include a weighted sum of the values ​​of each of the at least two radiomic features. The radiomic signature may be linearly related to the weighted sum of the values ​​of each of the at least two radiomic features.

[0083] The medical imaging data may include an attenuation value for each voxel of a plurality of voxels corresponding to at least the perivascular region. The plurality of voxels may also correspond to a blood vessel around which the perivascular region is disposed.

[0084] The method may further include using the medical imaging data to identify the perivascular region (or the substance, tissue or type of perivascular tissue, such as adipose tissue). The perivascular region can be identified as all voxels of the medical imaging data whose attenuation (or radiodensity) values ​​fall within a given range of attenuation values ​​and / or are positioned within a given radial distance from the outer vessel wall (i.e., the vessel is adjacent to the perivascular tissue). For example, if the perivascular region corresponds to perivascular adipose tissue, the given range of attenuation values ​​can be about -190 to about -30 Hounsfield units. For example, if the perivascular region corresponds to water, the given range of attenuation values ​​can be about -30 to about +30 Hounsfield units. The given radial distance can be a distance related to one or more dimensions of an adjacent blood vessel. The given radial distance can be equal to the adjacent blood vessel diameter. The given radial distance can be a fixed value, such as 5 mm.

[0085] The perivascular region may be adjacent to or surround a coronary artery, a carotid artery, aorta, or any other artery in a human body. For example, the perivascular region may be adjacent to or surround a proximal right coronary artery and / or a middle right coronary artery.

[0086] The method may further include segmenting the perivascular region. The method may further include calculating a value of the radiomics feature from the segmented perivascular region.

[0087] The value of each of the at least two radiomic features may be calculated based on raw attenuation values, binned attenuation values, or a wavelet transform of attenuation values.

[0088] The method may further comprise predicting the risk of the subject experiencing a major adverse cardiac event based on the calculated value of the radiomic signature. The method may further comprise predicting the risk of the subject experiencing a cardiac-specific major adverse cardiac event based on the calculated value of the radiomic signature.

[0089] The method may further include determining whether the subject suffers from a vascular disease based on the calculated value of the radiomic signature. The vascular disease may be selected from the group consisting of atherosclerosis, vascular calcification, intimal hyperplasia, vascular aneurysm, and vascular inflammation.

[0090] The method may further include determining whether the subject suffers from coronary heart disease based on the calculated value of the radiomics signature.

[0091] The calculated value of the radiomic signature may be used to distinguish unstable coronary artery lesions from stable coronary artery lesions.The perivascular region may be a region surrounding a lesion.

[0092] According to a second aspect of the present invention, a method for deriving a radiomic signature for predicting cardiovascular risk is provided. The method may include using a data set, such as a radiomics data set, to construct a perivascular radiomic signature or score for predicting cardiovascular risk. The radiomic signature may be calculated based on at least two radiomic features of the perivascular region. The data set may include measured values ​​of multiple radiomic features of the perivascular region, the measured values ​​of the multiple radiomic features being obtained from medical imaging data for each of a plurality of individuals. The plurality of individuals may include a first group of individuals that have reached a clinical endpoint indicative of cardiovascular risk and a second group of individuals that have not yet reached a clinical endpoint indicative of cardiovascular risk.

[0093] During a subsequent time period after the medical imaging data was collected, the first group of individuals may have reached a clinical endpoint indicative of cardiovascular risk, and during a subsequent time period after the medical imaging data was collected, the second group of individuals may not have reached a clinical endpoint indicative of cardiovascular risk.

[0094] Each of the at least two radiomic features can be selected to be collinear (or correlated) with a corresponding partner radiomic feature that is significantly (i.e., statistically significantly) associated with a clinical endpoint, e.g., as determined or calculated from a data set. The partner radiomic features of the at least two radiomic features can be different from each other.

[0095] Each of the at least two radiomic features may be selected as a different significant radiomic feature or co-linear with the different significant radiomic features, the different significant radiomic features being significantly associated with the clinical endpoint (as determined from the data set, i.e., identified from the data set as being significantly associated with the clinical endpoint). Thus, the method may include identifying significant radiomic features from a plurality of radiomic features that are each significantly associated with the clinical endpoint, the at least two radiomic features each being selected as a different significant radiomic feature or co-linear with the different significant radiomic features.

[0096] The important radiomic features may be selected to be non-collinear with each other. For example, the method may further include identifying a subset of multiple radiomic features that are non-collinear with each other (e.g., a subset of important radiomic features), at least two radiomic features are each selected to belong to a different radiomic feature of the subset or to be collinear with the different radiomic features.

[0097] Each of the partner radiomic signatures may be selected to be non-collinear with any of the other partner radiomic signatures.

[0098] At least one of the at least two radiomic signatures may be its own partner radiomic signature.

[0099] Each of the at least two radiomic features can be selected as being significantly associated with the clinical endpoint, eg, as determined or calculated from the data set.

[0100] The method may include identifying multiple clusters of radiomic features. Each cluster may include a subset of the multiple radiomic features. Each cluster may include the original radiomic feature, and each of the other radiomic features in the cluster is selected to be collinear, for example, as determined or calculated based on the data set. At least two radiomic features may each be selected from different clusters.

[0101] Each of the raw radiomic features may be selected to be non-collinear with any raw radiomic features of raw radiomic features of any other cluster in other clusters, for example, as determined or calculated from the data set.

[0102] Each of the radiomic features in each cluster can be selected to be collinear with all other radiomic features in the same cluster, for example, as determined or calculated from the data set.

[0103] Each of the raw radiomic features can be selected to be significantly associated with a clinical endpoint, eg, as determined or calculated from the data set.

[0104] Each of the raw radiomic signatures can be selected as being most closely associated with the clinical endpoint of all radiomic signatures in its cluster, eg, as determined or calculated from the dataset.

[0105] At least two radiomic features may be selected to be non-collinear with each other, eg, as determined or calculated from the data set.

[0106] Two radiomic features can be identified as collinear if their correlation is greater than a correlation threshold.

[0107] The Spearman rho coefficient can be used to calculate collinearity between radiomic features. Alternatively, other measures of pairwise correlation, such as the Pearson correlation coefficient (Pearson's r), can be used to calculate collinearity between radiomic features.

[0108] The correlation threshold may be at least about |0.75|, for example, about |rho|=0.75.

[0109] A radiomics feature may be determined to be significantly associated with a clinical endpoint if it is associated with the clinical endpoint above a significance threshold, e.g., as determined or calculated from the dataset. The significance threshold may be at least about α=0.05, e.g., about α=0.05.

[0110] As will be readily appreciated by those skilled in the art, associations of radiomic features with clinical endpoints can be calculated based on receiver operating characteristic (ROC) curve analysis, particularly using the area under the curve (AUC) measure (ie, C-statistic).

[0111] A Bonferroni correction can be applied to the significance threshold.

[0112] A principal component analysis of multiple radiomic features can be performed, for example, using a data set (specifically, based on the values ​​of the radiomic features of multiple individuals). The Bonferroni correction can be based on the number of principal components that take into account a given number of observed changes, as determined based on the principal component analysis. The given number of observed changes can be at least about 99.5%, for example, about 99.5%.

[0113] A radiomics signature can be constructed to be correlated with a clinical endpoint. A radiomics signature can be constructed to be significantly correlated with a clinical endpoint.

[0114] If the radiomic signature is associated with the clinical endpoint above a significance threshold, the radiomic feature can be identified as significantly associated with the clinical endpoint, e.g., as determined or calculated from the dataset. The significance threshold can be at least about 0.05, e.g., about α=0.05.

[0115] The data set may be divided into data for a training cohort of individuals and data for a validation cohort of individuals. The step of constructing a radiomics signature may include deriving a signature using data for at least the training cohort. The step of constructing a radiomics signature may include validating the signature using data for the validation cohort.

[0116] The radiomic signature can be identified as significantly associated with the clinical endpoint only if it is significantly associated with the clinical endpoint in both cohorts. If the radiomic signature is significantly associated with the clinical endpoint in the training cohort, the radiomic signature can be identified as significantly associated with the clinical endpoint.

[0117] Each of the at least two radiomic features may be independent of volume. Each of the at least two radiomic features may be independent of orientation. Multiple radiomic features may each be independent of volume and / or orientation.

[0118] Prior to selecting the at least two radiomic features, any volume- and / or orientation-related radiomic features may be removed from the plurality of radiomic features.

[0119] At least two radiomic features may be selected as being stable, eg, as determined or calculated from the data set. For example, at least two radiomic features may be selected from features identified as being stable, eg, as determined or calculated from the data set.

[0120] All unstable features may be removed from the plurality of radiomic features. For example, before selecting at least two radiomic features, all unstable features may be removed from the plurality of radiomic features.

[0121] If the intraclass correlation coefficient (ICC) for the radiomic feature (calculated for repeated measurements or scans) is less than a stability threshold, the radiomic feature can be identified as unstable. The stability threshold can be at least about 0.9, for example, 0.9. The intraclass correlation coefficient can be calculated based on the Z-score (or standard score, i.e., expressed as the number of standard deviations from the mean) of the radiomic feature.

[0122] The step of constructing a radiomic signature may include refining the contribution of at least two radiomic features to the radiomic signature to increase the association or correlation of the radiomic signature with the clinical endpoint. The radiomic signature may be constructed to be significantly correlated with the clinical endpoint.

[0123] As will be readily appreciated by those skilled in the art, association of a radiomics signature with a clinical endpoint can be calculated based on a receiver operating characteristic (ROC) curve analysis, particularly using the area under the curve (AUC) measure (ie, C-statistic).

[0124] The step of constructing the radiomics signature can be performed using a machine learning algorithm.

[0125] The step of constructing a radiomics signature can be performed using leave-out-p internal cross-validation.

[0126] The step of constructing the radiomics signature can be performed using leave-one-out internal cross-validation.

[0127] The step of constructing the radiomics signature can be performed using elastic net regression.

[0128] The radiomics signature may include a weighted sum of at least two radiomics features.

[0129] The radiomic signature may be linearly related to a weighted sum of at least two radiomic features.

[0130] The step of constructing a radiomic signature may include adjusting the relative weight of each of the at least two radiomic features to increase the association or correlation of the radiomic signature with the clinical endpoint.

[0131] Radiomic signatures can be constructed to provide a measure of the texture of the perivascular region.

[0132] At least one of the at least two radiomic features may provide a measure of the texture of the perivascular region. For example, each of the at least two radiomic features may provide a measure of the texture of the perivascular region (i.e., each of the at least two radiomic features may be a texture statistic).

[0133] The clinical endpoint may be a composite endpoint of major adverse cardiac events. The clinical endpoint may be a composite endpoint of cardiac-specific major adverse cardiac events.

[0134] The method of the present invention may further comprise the step of calculating radiomics features from the medical imaging data.

[0135] In addition to the at least two radiomic features mentioned above, the radiomic signature of the present invention can also be calculated based on additional radiomic features. Therefore, it can be said that the radiomic signature is calculated based on multiple radiomic features, and multiple radiomic features can include at least two radiomic features. For example, a method for deriving a radiomic signature for predicting cardiovascular risk can include using a data set, in particular a radiomic data set, to construct a perivascular radiomic signature or score for predicting cardiovascular risk. The radiomic signature can be calculated based on (a second) plurality of perivascular radiomic features (i.e., radiomic features of the perivascular region). The data set can include measured values ​​of (a first) plurality of perivascular radiomic features of the perivascular region, and the measured values ​​of the (first) plurality of perivascular radiomic features are obtained from medical imaging data for each individual in a plurality of individuals. The plurality of individuals can include a first group of individuals that have reached a clinical endpoint indicating cardiovascular risk and a second group of individuals that have not yet reached a clinical endpoint indicating cardiovascular risk. The second plurality of perivascular radiomic features may be selected from the first plurality of perivascular radiomic features, in particular to provide a radiomic signature for predicting cardiovascular risk, as determined based on or using the dataset, e.g., using a machine learning algorithm. Thus, in addition to the at least two radiomic features, a radiomic signature may be calculated based on further radiomic features (e.g., selected from the (first) plurality of radiomic features).

[0136] The method may further include configuring a system for calculating the value of the patient's radiomic signature. For example, the method may further include configuring the system to characterize the perivascular region of the patient or subject by calculating the value of the derived radiomic signature of the patient or subject. The system may be configured to calculate the value of the derived radiomic signature using or based on medical imaging data of at least the perivascular region of the patient or subject. The system may be configured to calculate the value of the derived radiomic signature using or based on at least the values ​​of at least two (or a second plurality) of radiomic features of the perivascular region of the patient or subject.

[0137] Thus, methods may be used to derive a perivascular radiomic signature and configure a system for characterizing the perivascular region of a patient using the derived radiomic signature.

[0138] The system may be configured to receive as input medical imaging data or values ​​of at least two (or a second plurality) of radiomic features. The system may be configured to output (e.g., display) a calculated value of a radiomic label or output a value based on the calculated value of a radiomic label. The system may be configured to output an indication of the patient's cardiovascular risk and / or vascular health. The system may be configured to output an indication of the patient's risk of experiencing an adverse cardiovascular event. The system may be a computer system.

[0139] The method may include providing instructions for configuring a system for calculating a value of a derived radiomic signature for a patient or subject.

[0140] The method may further include calculating the value of the derived radiomic signature for the perivascular region of the patient or subject. For example, the method may further include characterizing the perivascular region of the patient or subject by calculating the value of the derived radiomic signature. The value of the derived radiomic signature may be calculated based on or using medical imaging data of at least the perivascular 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 at least two (or a second plurality) of radiomic features of the perivascular region of the patient or subject.

[0141] Therefore, the method can be used to derive a perivascular radiomic signature and use the derived radiomic signature to characterize the perivascular region.

[0142] According to a third aspect of the invention, there is provided a system configured to perform any of the methods described above. 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.

[0143] According to a fourth aspect of the present invention, there is provided a use of a perivascular radiomic signature for any of the above purposes, for example, for characterizing a perivascular region, detecting vascular disease or predicting cardiovascular risk. The perivascular radiomic signature can be calculated based on measured values ​​of a plurality of perivascular radiomic features of a perivascular region.

[0144] The medical imaging data may be radiographic data. The medical imaging data may be computed tomography data.

[0145] The perivascular region may be or may include perivascular tissue, for example, perivascular adipose tissue. The perivascular region may also include water and / or other soft tissue structures within the perivascular region. BRIEF DESCRIPTION OF THE DRAWINGS

[0146] The present invention will now be described with reference to the accompanying drawings, in which:

[0147] Figure 1 shows the research methodology and design. Figure 1AMethods for perivascular region phenotyping, volume segmentation, discretization, wavelet transform, and calculation of radiomic features are presented. Figures 1B and 1C show a summary of cohorts 1 and 2 of the study, respectively. AMI: acute myocardial infarction; CAD: coronary artery disease; CCTA: coronary computed tomography angiography; MACE: major adverse cardiovascular event; PVR: perivascular region; H / L: high / low wavelet transform; VOI: volume of interest.

[0148] Figure 2 summarizes the construction of the radiomic signature of high-risk perivascular regions. Figure 2A Summarize the multi-step method for labeling as a flow chart.First, in a total of 843 calculated radiomic features, 124 were excluded after stability analysis (ICC<0.9).Next, sixty radiomic features significantly associated with MACE at a=0.05 level in two queues were selected.Among these, seven features related to volume and orientation (e.g., energy and non-HHH / LLL transformation) were excluded, and then by gradually eliminating the paired correlation at the level of |rho|≥0.75, collinearity was reduced.Using the remaining nine radiomic features, a machine learning method using elastic network regression and leave-one-out internal cross-validation was applied in queue 1. Figure 2B The relative variable importance (weighting) of the best performing models for prediction of MACE is shown. Figure 2C The relative variable importance (weighted) of the best performing models for prediction of cardiac-specific MACE (cMACE) is shown. Figure 2C The first six components of the model are used to define the perivascular texture index (PTI), which is a radiomics signature according to the present invention. AUC: area under the curve; CI: confidence interval; MACE: major adverse cardiovascular events; SALGLE: small area low grayscale advantage; SRHGLE: short run high grayscale advantage; SZNU: size zone heterogeneity.

[0149] Figure 3 shows that the PTI signature of the present invention provides incremental prognostic information beyond current CCTA-based tools. In a pooled analysis of the two Group 1 cohorts, the perivascular texture index (PTI) was added to the high-risk signature derived from age, sex, traditional risk factors, and computed tomography (including pericoronary FAI). PVAT ) significantly improved the predictive value of the model for both MACE (δ[ΑUC]=0.020, P=0.014, Figure 3A) and cardiac-specific MACE (δ[ΑUC]=0.046, P=0.012, Figure 3B). PVATThe risk of both endpoints was independently associated with higher values ​​of both biomarkers, with higher prospective risk of adverse events. Figure 3C The interaction between two CT-derived perivascular region (PVR) features (PTI and FAI) is presented graphically using bidirectional contour plots depicting the adjusted probability of MACE (left) and cardiac-specific MACE (right) across different levels of the two biomarkers. These findings suggest that FAI PVAT Both CT and PTI have significant clinical value in comprehensively phenotyping the perivascular area of ​​the coronary artery.

[0150] Figure 4 Correlation plots of raw radiomic features are shown. Correlation plots of 107 coronary perivascular regional radiomic features are shown, calculated from raw CCTA images. The plots depict the direction and strength of association between individual radiomic features (calculated by Spearman's rho coefficient) and reveal the presence of distinct feature clusters.

[0151] Figure 5 shows various aspects of principal component analysis of the radiomic features of the perivascular regions of the coronary vessels. Figure 5A The component plots of the three main principal components in Group 1 are shown. Figure 5B A scree plot is shown depicting the percentage of variation explained by the 92 first components, which account for 99.5% of the variation. Figure 5C Manhattan plots of the 5-year MACEs for the 42 components with predicted eigenvalues ​​> 1 are shown. Figure 5D Shown are forest plots presenting adjusted odds ratios (statistically adjusted for covariates such as age, sex, etc.) for predicting 5-year MACE for seven principal components significant in univariate analysis. Figure 5E Correlation plots of selected components with demographics are shown. Fig. 5F The incremental prognostic value of the perivascular area radiomic components for the prediction of 5-year MACE is presented. AUC: area under the curve; BMI: body mass index; CCS: coronary artery calcium score; DM: diabetes mellitus; EAT: epicardial adipose tissue; HCH: hypercholesterolemia; HDL: high-density lipoprotein; LDL: low-density lipoprotein; PC: principal component; PVR: perivascular area; SBP: systolic blood pressure; TG: triglycerides.

[0152] Figure 6 Manhattan plots for predicting 5-year MACE are shown. For all radiomic features calculated on original or wavelet transformed images, P values ​​were derived from receiver operating characteristic (ROC) curve analysis for discriminating 5-year MACE. MACE: major adverse cardiovascular events; H / L: high-pass / low-pass signal in wavelet transform.

[0153] Figure 7 A plot of the intraclass correlation coefficient (ICC) of all perivascular region radiomic features is shown. In repeated analyses, radiomic features were ranked according to the intraclass correlation coefficient (ICC). A total of 719 radiomic features were found to have an ICC ≥ 0.9.

[0154] FIG8 shows that perivascular area texture phenotyping using a radiomics tag according to the present invention can be used to detect unstable coronary lesions. FIG8A-G shows a box line drawing of the six radiomics features included in PTI and the PTI itself. In Group 2, five of the six component radiomics features of PTI were significantly altered in the presence of unstable coronary lesions versus stable coronary lesions, and PTI outperformed all six individual radiomics features in distinguishing unstable forms of stable coronary lesions (AUC: 0.76; 95% CI: 0.64-0.88). These findings suggest a close link between vascular inflammation associated with plaque rupture and changes in perivascular area texture (which can now be detected by the radiomics tags of the present invention, e.g., PTI) as a noninvasive way to monitor the risk profile of coronary lesions. AU: arbitrary unit; SALGLE: small area low grayscale advantage; SRHGLE: short run high grayscale advantage; SZNU: size zone heterogeneity. DETAILED DESCRIPTION

[0155] The inventors have discovered that a perivascular region (PVR) radiomic signature (otherwise referred to as a "score" or "index") calculated based on two or more radiomic features of PVR adds incremental value beyond traditional risk factors and established CCTA risk classification tools in predicting future adverse cardiovascular events and assessing cardiovascular health and risk, and further aids in detecting general vascular inflammation, localized plaque inflammation, and the presence of unstable coronary lesions. Therefore, the PVR radiomic signature of the present invention is preferably calculated based on two or more radiomic features of PVR, and provides a tool for characterizing PVR (e.g., perivascular tissue, such as perivascular adipose tissue (PVAT)) for assessing cardiovascular or vascular health, predicting a patient's risk of future adverse cardiovascular events, identifying or diagnosing coronary artery disease or coronary heart disease, and identifying unstable coronary lesions or vascular inflammation (e.g., caused by localized plaques).

[0156] The PVR radiomics signature of the present invention can be used alone to characterize PVR or provide diagnostic or prognostic information, or can be combined with existing models, such as FAI PVAT, Duke Prognostic CAD Index, and / or other conventional models incorporating demographic characteristics and risk factors, such as the presence of coronary artery calcium, high-risk plaque characteristics, and / or EAT (epicardial adipose tissue) volume.

[0157] The present invention exploits the fact that the coronary wall and adjacent PVR, particularly tissue within the PVR, such as adipose tissue, interact in a bidirectional manner. Therefore, vascular-induced phenotypic changes in coronary PVR can act as sensors for underlying disease even in the absence of visible coronary artery lesions. Specifically, the present invention exploits the effects of this interaction on the texture of the PVR (e.g., spatial heterogeneity or variability), and thus the radiomic signature of the present invention can be constructed to provide a measure of the texture of the PVR. Therefore, the radiomic signature of the present invention can also be referred to as the perivascular texture index (PTI). However, the radiomics-based method used to construct the signature of the present invention is not specific to constructing a radiomic signature that measures texture, and what is most important is the prognostic value of the resulting signature. Therefore, it is not necessary for a radiomic signature to measure texture in order to be an effective prognostic or diagnostic tool.

[0158] PVR refers to the area or volume adjacent to a blood vessel. PVR may be an area or volume of perivascular tissue (PVT) or may include or consist of PVT. Perivascular tissue is tissue located near a blood vessel. Tissue is a complex biological structure and may include cells (e.g., adipocytes, neurons, etc.) as well as extracellular structures and materials (such as water) that may occupy the intercellular spaces. Specifically, PVT may include or consist of perivascular adipose tissue (PVAT), and thus PVR may alternatively be referred to as an area or volume of PVAT.

[0159] The present invention employs a radiomics approach. Radiomics is an imaging field in which data characterization algorithms are used to extract large amounts of quantitative information from imaging data. The resulting features, referred to as radiomic features, range from simple volumetric, shape-related, or first-order statistics such as mean or median attenuation to second- and higher-order statistics describing the texture of segmented volumes or regions and the spatial relationships of voxels with similar or different attenuation values. Such features can identify imaging patterns of significant clinical value that are not discernible to the naked eye and have the potential to maximize the diagnostic yield of noninvasive PVR phenotyping.

[0160] The labels of the present invention are 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 corresponds to a perivascular region (PVR), for example, a coronary PVR, such as coronary perivascular adipose tissue (PVAT), and optionally also corresponds to the blood vessel itself and / or other tissues adjacent to or surrounding the PVR. Medical imaging data typically includes radiodensity (or attenuation) values ​​for multiple voxels of the relevant region (in this case, the PVR and optionally also adjacent tissues), typically expressed in Hounsfield units (HU).

[0161] The medical imaging data is preferably computed tomography (CT) data, such as coronary computed tomography angiography (CCTA), 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 CCTA or another medical imaging technique is mentioned, it should be understood that other suitable medical imaging techniques may be used alternatively.

[0162] The PVR may include only voxels whose radiodensity (or attenuation) falls within a given or predetermined range and / or is positioned within a defined region, for example, within a given or predetermined radial distance from the outer vessel wall. The given radial distance is preferably a distance related to or dependent on one or more dimensions of an adjacent vessel (such as its diameter or radius). However, the radial distance may alternatively be a set value or fixed value, such as about 5 mm. Alternatively, the PVR may be identified by a human operator through manual contouring or delimitation, optionally in combination with applying a radiodensity or attenuation mask, so that only voxels whose radiodensity is within a specified range and falls within the defined region are included. For example, an operator may identify the PVR by examining data (e.g., a CT image). The PVR may include only voxels whose radiodensity is within a range of about -190HU to about +30HU Hounsfield units. For example, the PVR may include only voxels whose radiodensity is within a range of about -190HU to about -30HU Hounsfield units. This range of attenuation values ​​typically corresponds to the radiodensity of perivascular adipose tissue (PVAT). However, other ranges may be used, such as, for example, about -30 to about +30 Hounsfield units, which generally corresponds to the radiodensity of water. Specifically, PVR may be identified as all voxels having a radiodensity in the range of about -190 HU to about -30 HU Hounsfield units and located within a radial distance from the adjacent outer vessel wall that is approximately equal to the adjacent vessel diameter.

[0163] Radiomic features and thus radiomic labels can be calculated for specific blood vessels, for example, coronary blood vessels, such as coronary arteries. Due to their straight course and the absence of large branches, the proximal right coronary artery (RCA) and the middle right coronary artery (segment 1 and segment 2 according to the anatomical classification of the American Heart Association, such as defined in Austen WG, Edwards JE, Frye RL et al., in the reporting system for evaluating coronary artery disease in patients, the Ad Hoc Committee for Grading of Coronary Artery Disease, Council on Cardiovascular Surgery, American Heart Association, issued 1975; 51 (4 Supplement): 5-40) are particularly suitable. Therefore, PVR can be positioned near a specific blood vessel.

[0164] The PVR may be segmented prior to calculating radiomic features, and the radiomic features may be calculated from the segmented data. The segmented volume or region corresponds to the PVR, and the segmentation may remove data corresponding to voxels outside the PVR. Thus, segmentation may be achieved by identifying the PVR as described above, and then removing any voxels from the data identified as not being part of the PVR, e.g., those corresponding to surrounding or adjacent tissue voxels. Segmentation may be performed by placing a three-dimensional sphere having a diameter equal to the vessel diameter plus twice a given distance from the outer vessel wall within which the PVR may be identified on consecutive slices along the vessel centerline. For example, if the PVR is identified as being positioned within a radial distance from the outer vessel wall equal to the adjacent vessel diameter, the diameter of the sphere would be equal to three times the adjacent vessel diameter. The segmented PVR may be extracted and used to calculate radiomic features.

[0165] The calculation of radiomic features from medical imaging data can be performed using a computer program or software. There are various commercially available software packages for this purpose, such as 3D Slicer. The radiomic features can be shape-related statistics, first-order statistics, or texture statistics (e.g., second-order and higher-order statistics). Shape-related radiomic features and first-order radiomic features can be calculated using the raw radiodensity (HU) values ​​of the PVR voxels. For the calculation of texture features (e.g., gray-level co-occurrence matrix [GLCM], gray-level dependence matrix [GLDM], gray-level run length matrix [GLRLM], gray-level size zone matrix [GLSZM], and neighboring gray-level tone difference matrix [NGTDM], see Figure 1A and Tables R1-R7), the PVR voxel radiodensity or attenuation values ​​are preferably discretized into multiple bins, preferably into 16 bins, preferably of equal width (e.g., a width of ten HU), to reduce noise while allowing sufficient resolution to detect biologically significant spatial variations in PVR 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. In order 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 four directions).

[0166] Some or all of the radiomic features, particularly those related to first-order statistics and texture-based statistics, may also be calculated for the three-dimensional wavelet transform of the raw image data, thereby generating a plurality of additional radiomic feature sets ( Figure 1A), 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 original 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 original 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 passing the data through a multi-level (e.g., three-level) filter bank through a discrete wavelet transform. At each level, the data is decomposed into high-frequency 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 data. Such decomposition can be performed using widely available software (such as the Slicer Radiomics software package incorporated into the Pyradiomics library). In the case of calculating radiomic features based on wavelet decomposition or transformation of data, this is represented by the suffix of the radiomic features calculated based on wavelet decomposition (for example, HHH represents high pass, high pass, high pass). Therefore, for example, "skewness LLL" represents the radiomic feature "skewness" calculated as based on LLL wavelet decomposition. In the absence of a suffix, the radiomic features are calculated based on original (or raw) data.

[0167] Deriving radiomic signatures

[0168] The present invention provides a method for deriving a radiomics signature for characterizing PVR (e.g., area of ​​perivascular adipose tissue) for example for predicting cardiovascular risk. The radiomics signature is derived using medical imaging data for a plurality of individuals and data containing the occurrence of a clinical endpoint event for each of the plurality of individuals in a subsequent time period after the medical imaging data was collected. Specifically, the clinical endpoint is preferably indicative of cardiovascular health or risk.

[0169] The method generally involves conducting a case-control study of (human) patients with and without adverse (clinical endpoint) events within a predetermined or subsequent time period, preferably within five years, after collecting imaging data by CCTA or other medical imaging techniques, for example, by clinically indicated assessment. Individuals who reach the clinical endpoint are cases (first group), and individuals who do not reach the clinical endpoint are controls (second group).

[0170] Multiple individuals (also referred to herein as patients) can be divided into two independent queues of medical imaging that patients are undergoing (or described patients have experienced), specifically training queue and verification queue. Based on the occurrence of clinical endpoint events in a specified or predetermined time period (i.e., subsequent time period) after collecting medical imaging data to identify cases. The subsequent time period is preferably at least about five years, preferably about five years, but can be longer or shorter time. The clinical endpoint is preferably the main composite endpoint of major adverse cardiovascular events (MACE), which can be defined as the composite of all-cause mortality and non-fatal myocardial infarction (MI) in a specified time period (preferably five years) after collecting medical imaging data, or the main composite endpoint of cardiac specific MACE (cMACE, i.e., cardiac mortality and non-fatal MI) in a specified time period (preferably five years) after collecting medical imaging data.

[0171] As used herein, cardiac mortality and non-cardiac mortality can be defined according to the ACC / AHA recommendations (Hicks KA, Tcheng JE, Bozkurt B, et al., 2014 ACC / AHA Key Data Elements and Definitions for Cardiovascular Endpoint Events in Clinical Trials: A Report of the American College of Cardiology / American Heart Association Task Force on Clinical Data Standards (Writing Committee for Developing Cardiovascular Endpoint Data Standards), J Am Coll Cardiol 2015; 66(4): 403-69). More specifically, cardiac mortality can be defined as any death due to adjacent cardiac causes (e.g., myocardial infarction, low-output heart failure, fatal arrhythmias). Deaths that meet the criteria for sudden cardiac death can also be included in this group. Any death not covered by the previous definition, such as death due to malignancy, accident, infection, sepsis, renal failure, suicide, or other noncardiovascular causes (such as stroke or pulmonary embolism), can be classified as noncardiac.

[0172] Controls are preferably identified as patients who have been followed up for event-free events within the same specified time period, e.g., for at least five years after CCTA. Cases are preferably matched to controls using, for example, an automated algorithm, 1:1 case-control matching is preferably performed to match cases to controls. Cases and controls may be matched for clinical demographic characteristics (e.g., age, sex, obesity status), cohorts, and / or technical parameters related to imaging data acquisition (e.g., tube voltage and CT scanner used). Patients are preferably also matched for other cardiovascular risk factors, including hypertension, dyslipidemia, diabetes, and smoking.

[0173] Hypertension was defined based on the presence of a documented antihypertensive regimen or treatment according to relevant clinical guidelines (James PA, Oparil S, Carter BL, et al., 2014 evidence-based guideline for the management of high blood pressure in adults: report from the panel members appointed to the Eighth Joint National Committee (JNC 8), JAMA 2014;311(5):507-20). Similar criteria can be applied to the definitions of hypercholesterolemia and diabetes (American Diabetes A. Diagnosis and classification of diabetes mellitus. Diabetes Care 2014;37 Suppl 1:S81-90; Stone NJ, Robinson JG, Lichtenstein AH, et al., 2013 ACC / AHA guideline on the treatment of blood cholesterol to reduce atherosclerotic cardiovascular risk in adults: a report of the American College of Cardiology / American Heart Association Task Force on Practice Guidelines. J Am Coll Cardiol 2014;63(25Pt B):2889-934).

[0174] "Clustering" (or collinearity elimination) methods

[0175] A radiomic signature may then be developed in a stepwise approach, as schematically illustrated in FIG2 . First, for example, as described above, a plurality of 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 the original image data and / or one or more wavelet transforms (or wavelet decompositions) of the image data, as described above. Preferably, each of the radiomic features is calculated for the original image data and for the above-described eight three-dimensional wavelet decompositions of the image data.

[0176] Unstable features can be removed from multiple radiomic features. A z-score transformation can be applied to the feature (i.e., the value of the radiomic feature is expressed according to the number of standard deviations relative to the mean), and a stability analysis can be performed based on the z-score. Unstable radiomic features are identified as unstable radiomic features with an intraclass correlation coefficient (ICC) below a stability threshold in repeated imaging data acquisition (e.g., imaging scans). For example, the stability threshold can be at least about 0.9, for example, about 0.9, so that all radiomic features with ICC<0.9 are excluded (Figure 3). However, other stability thresholds, such as 0.85 or 0.95, can be used instead. ICC can be calculated for multiple repeated scans, for example, two to ten scans, specifically two or ten scans. In other words, stability analysis can be performed on PVR radiomic features and unstable radiomic features removed from multiple radiomic features. Stability analysis can be performed based on imaging data for multiple individuals, or can be performed using other data, for example, reference data, such as the RIDER dataset (reference image database for evaluating treatment response).

[0177] Preferably, if the plurality of radiomic features includes any volume- and orientation-related radiomic features, they are excluded (i.e., removed from the plurality of radiomic features). Alternatively, any volume- and orientation-related features may be excluded from the outset, so that this step is not required. In other words, the initial plurality of radiomic features may each be volume- and orientation-independent. Volume-related features may include energy and total energy (raw and wavelet calculated), and orientation-related features may include features derived from wavelet transforms HLL, HLH, LLH, HLL, LHL, LHH (i.e., all features except LLL or HHH, or those wavelet transforms that are not limited to high-pass or low-pass).

[0178] Radiomic features that are not significantly associated (e.g., correlated) with clinical endpoints can then be removed from multiple radiomic features, for example, 5-year MACE or cMACE above a significance threshold. The association of each radiomic feature with a clinical endpoint can be calculated based on a receiver operating characteristic (ROC) curve analysis, specifically calculating the area under the curve based on data from multiple individuals. The significance threshold is preferably 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. Preferably, the radiomics feature must be identified as being significantly associated with the clinical endpoint in both the training cohort and the validation cohort in order to retain it, and remove any radiomics feature that is found not to be significantly associated with the clinical endpoint in one or both cohorts to reduce the positive findings due to cohort-specific changes. However, only the radiomics features that are found to be significantly associated with the clinical endpoint in one of the cohorts (e.g., the training cohort) can be retained instead. Alternatively, an assessment of whether each radiomics feature is significantly associated with the clinical endpoint can be performed based on summary data for both cohorts, i.e., for all individuals in a plurality of individuals. In other words, the method may involve assessing the significance of the radiomics feature with the clinical endpoint, and removing features that are found not to be significantly associated with the clinical endpoint from a plurality of features. The final result should be to remove any radiomics feature that is not significantly associated with the clinical endpoint from a plurality of radiomics features (as determined or calculated according to the data, e.g., based on data analysis).

[0179] The collinearity of the retained radiomic features (i.e., radiomic features significantly associated with clinical endpoints, also referred to as important radiomic features) can then be reduced or eliminated by removing pairwise correlations, i.e., by removing at least one of each pair of collinear radiomic features. The correlations between radiomic features are typically calculated using measured values ​​of the radiomic features of multiple individuals. The removal of pairwise correlations can be performed in a stepwise manner. Collinear radiomic features can be identified as radiomic features that are correlated to each other to a degree at least equal to a given correlation threshold. 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 pairwise correlations at the level of |rho|≥0.75 are eliminated. 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.

[0180] For example, when a pair of collinear radiomic features are identified, one of the two features is preferably eliminated from a plurality of features. For example, the radiomic features that are not so closely associated with the clinical endpoints of both calculated according to the data can be eliminated, and the radiomic features that are most closely associated with the clinical endpoints can be retained, but this is not necessary, and any one can be retained or deleted. For example, the collinear elimination step can be performed in an unsupervised manner without considering the clinical endpoint, and the algorithm can eliminate the maximum redundant features that contribute the least to the change of the research group (for example, features with smaller variances such as those measured across multiple individuals). In an example, when a pair of collinear features are identified, the feature with the maximum average (for example, mean) absolute correlation (i.e., the average correlation value (or average modulus or square correlation value) with all other radiomic features) is removed. This can be performed in a stepwise manner until the collinear radiomic features are not retained.

[0181] The collinear elimination step can be performed using an algorithm or function (e.g., the function claret:findCorrelation, R package (see Kuhn, M. & Johnson, K, Applied Predictive Modelling, (Springer, 2013)). For example, the function or algorithm can construct a pairwise correlation matrix containing pairwise correlations between radiomic features. The function can then search through the correlation matrix and return an integer vector corresponding to the columns to be removed to reduce the pairwise correlations. The radiomic features corresponding to these columns can then be removed from the plurality of radiomic features. In determining which columns to remove, the algorithm can first identify the pairwise correlations between the radiomic features. When two collinear radiomic features are identified, the algorithm then identifies the column corresponding to the feature with the largest average absolute correlation for removal.

[0182] Regardless of how the collinearity elimination step is performed, the end result is preferably a reduced plurality of radiomic features in which each of the features is less correlated with each of the other remaining features by less than a correlation threshold. In other words, the method may involve the step of removing radiomic features to eliminate collinearity between the radiomic features, such that none of the remaining radiomic features is collinear with any of the other remaining radiomic features. This may involve calculating pairwise correlations between the radiomic features, and removing at least one of any identified pair of collinear features.

[0183] A radiomics signature can then be constructed based on at least two of the remaining radiomics features that survived any of the steps described above (e.g., stability analysis, significance analysis, and / or collinearity elimination). For example, a radiomics signature can then be constructed based on at least two of the reduced multiple non-collinear radiomics features that survived the collinear elimination step. The reduced multiple features that survived the collinear elimination step are otherwise referred to as "original features." However, since the eliminated radiomics features are each closely related to at least one of the original features, a label in which one or more of the original features is replaced by one of the features that are collinear with the replaced original features will generally be performed in a manner similar to that of calculating a label based on only the original features. For example, it is possible to exchange one of the original features for one of the features that is calculated to be collinear with the original feature, and the label should be performed in a similar manner. In fact, it is possible that replacing one or more (or even all) of the original features with alternative features that are co-linear with the replaced original features may result in a signature with improved prognostic value, and in fact has been found to be so in some cases (see, e.g., Tables 8A and 8B). This is because although the original features are often most independently associated with clinical endpoints, they are not necessarily the best performing features when combined into a signature.

[0184] Thus, the process of constructing a radiomic signature can involve constructing "clusters" of radiomic features (each cluster including one of the original features), in which each of the radiomic features in each cluster is collinear with at least the "original" feature in the cluster (i.e., the feature of the cluster that survived the collinear elimination step, e.g., the feature in the cluster that is most independently closely associated with the clinical endpoint). The construction of these clusters can be performed in lieu of a collinear elimination step. For example, instead of eliminating one of each pair of collinear features, the collinear features can be assigned to the same cluster. Alternatively, a pairwise elimination step can be performed as described above, and then, once the original features are identified, the eliminated features can be reintroduced by assigning them to the cluster of the original features with which they are most closely associated or collinear.

[0185] However, no matter how the clusters are constructed, the end result should be that each radiomic feature and one or more original radiomic features that are collinear with it are assigned to the same cluster. If a radiomic feature is collinear with two "original" features, it is preferred that the radiomic feature be assigned to the cluster of the original feature that is most collinear with it, but it can also be assigned to the cluster of all original features that are collinear with it.

[0186] A cluster may also be expanded to include any of the original plurality of radiomic features that are collinear with the "original" radiomic feature of the cluster, regardless of whether the radiomic feature itself is independently significantly associated with the clinical endpoint. However, as previously mentioned, preferably any radiomic feature included in a cluster is stable.

[0187] Thus, the "original" radiomic feature in each cluster represents a "partner" radiomic feature of each of the other radiomic features in the cluster, where each of the radiomic features in each cluster is collinear with its "partner" feature. Thus, in this sense, the original radiomic feature can be considered to be its own "partner" radiomic feature, as it is completely collinear with itself.

[0188] The initial radiomic signature can then be constructed based on at least two (or all) of the reduced plurality of features (e.g., "original" features). Alternatively, if clusters are constructed, the initial radiomic signature can be constructed from at least two radiomic features, each selected from a different cluster. The construction of the radiomic signature involves refining or optimizing the radiomic signature, specifically using data from a "training" cohort. This involves refining or optimizing the contribution of each of the remaining radiomic features to the signature based on the data to improve the correlation or association between the signature and the clinical endpoint. For example, the signature can include a weighted sum of the values ​​of each radiomic feature in the radiomic features included in the initial signature, and the weight of each radiomic feature in the radiomic features can be gradually optimized or improved. The coefficient by which each radiomic feature in the radiomic features is multiplied is typically referred to as a beta (β) coefficient, and it is these β coefficients that can be optimized or improved.

[0189] Preferably, all reduced multiple features are included in the initial radiomic signature to be improved, for example, one feature from each cluster, but this is not absolutely necessary. For example, the "best" feature or "original" feature from each cluster (for example, the feature most independently closely associated with the clinical endpoint) can be included. Other radiomic features can also be included in the initial signature to be optimized, for example, two or more radiomic features from any or all of the clusters can be included in the initial signature. However, in order to provide a label that is more closely associated with the clinical endpoint 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 information related to PVR. Specifically, radiomic features from different clusters will be sensitive to different phenotypic characteristics of PVR because they are collinear with different "original" or "partner" features. For example, the initial radiomic signature can include at least three radiomic features, each of which is selected from a different cluster. Alternatively, the initial radiomic signature may include at least four radiomic features, each radiomic feature selected from a different cluster. Alternatively, the initial radiomic signature may include at least five radiomic features, each radiomic feature selected from a different cluster. Alternatively, the initial radiomic signature may include at least six radiomic features, each radiomic feature selected from a different cluster. Alternatively, the initial radiomic signature may include at least seven radiomic features, each radiomic feature selected from a different cluster. Alternatively, the initial radiomic signature may include at least eight radiomic features, each radiomic feature selected from a different cluster. Preferably, the initial radiomic signature may include one radiomic feature from each cluster.

[0190] Once the initial radiomic signature has been perfected, a final radiomic signature can be constructed based on all radiomic features or a subset of these features included in the initial signature. For example, the optimized or perfected initial signature can be used as the final signature. Alternatively, the final radiomic signature can include only a subset of the radiomic features included in the initial signature. For example, only those parameters whose contribution to the improved initial signature is above a certain threshold can be included in the final signature, such as those parameters whose optimized weighting coefficients (e.g., β parameters) are above a certain (e.g., predetermined) threshold. In other words, those radiomic features whose optimized coefficients are below a given threshold will be removed from the final signature. For example, the first seven features (i.e., the features that contribute the most to the perfected signature) can be included in the final signature.

[0191] If only a subset of radiomic features are included in the final signature, one option is to re-optimize or refine the signature based on only the subset of features. For example, coefficients or beta parameters can be re-optimized for the subset of radiomic features included in the final signature. Alternatively, coefficients or beta parameters derived from an initial optimization performed on the initial signature can be used in the final signature.

[0192] As mentioned above, a label may comprise a weighted sum of the calculated values ​​of multiple radiomic features. A label may also comprise other terms, such as addition or subtraction of a constant, or multiplication of a factor. However, typically, a label will be linearly related in some way to the weighted sum of the radiomic feature values.

[0193] Radiomics labels can be of the following forms or contain terms (e.g., labels can be calculated based on functions of containing terms):

[0194] A ± Σb i rf i

[0195] 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 the radiomic feature i. The β parameter in the equation is preferably an unnormalized β, which is the Z-score β parameter multiplied by the standard deviation of the radiomic feature, where the standard deviation is preferably the standard deviation of the radiomic feature in the training cohort or the derivation cohort. The constant A is not required, but can be included to ensure that all result values ​​are positive or negative, preferably positive. Preferably, "±" is "-".

[0196] Machine learning algorithms can be used to construct (i.e., optimize or improve) initial and / or final radiomics signatures. For example, a machine learning algorithm can be used to improve or optimize the contribution of each radiomics feature in the radiomics signature to the signature, for example by optimizing the β coefficient. The machine learning method can use elastic net / lasso regression, and can utilize leave-one-out internal cross-validation. Elastic net regression is a regularized regression method that linearly combines the L1 and L2 penalties of the lasso and ridge methods. The absolute value of the L1 (least absolute shrinkage and selection operator "lasso" penalty) penalty coefficient shrinks irrelevant coefficients to zero and helps feature selection, while the square of the L2 ("ridge" penalty) penalty coefficient, thereby limiting the impact of collinearity and reducing overfitting. The optimal penalty coefficient (λ, λ) can be selected by cross-validation, and α is preferably set to α=1.

[0197] Preferably, the labels are refined using the data for the training cohort, and then validated externally using the data for the training cohort.

[0198] Non-clustering method

[0199] The present invention also provides another method for deriving a radiomics signature. In this method, as described above with respect to the "clustering" method, a plurality of radiomics features are calculated for each of a plurality of individuals from medical imaging data.

[0200] Then, for example, receiver operating characteristic (ROC) curve analysis can be used, specifically using area under the curve calculation (calculating C statistics) to evaluate the discriminative value of all radiomic features (preferably all stable radiomic features) for clinical endpoints (e.g., 5-year MACE), and any radiomic features that are not significantly associated with the clinical endpoint are eliminated from multiple features. For the clustering method described above, if a feature is associated with a clinical endpoint above a certain significance threshold, the feature is identified as being significantly associated with the clinical endpoint. The significance threshold is preferably α=0.05 or less, for example, α=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 may be about α=0.002.

[0201] In order to correct for multiple comparisons and reduce the false discovery rate (FDR), the Bonferroni correction can be applied to the significance threshold. The Bonferroni correction can be applied based on the number of principal components, which accounts for a given number of variability in the research sample based on principal component analysis. For example, the given number can be about 99.5%. In other words, the m value for correcting the α value (by dividing α by m, that is, α / m) is the number of principal components that account for a given number of variability. Therefore, principal component analysis of radiomics features can be performed on the data of multiple individuals (if individuals are divided into training queues and verification queues, preferably two queues are included).

[0202] As described above, a radiomic signature can then be constructed based on at least two of the remaining radiomic features, i.e., radiomic features that are identified as being significantly associated with the clinical endpoint. A radiomic signature can be constructed and optimized in the same manner as described above for the clustering method, but with the difference that radiomic features do not need to be selected from different clusters.

[0203] For example, all radiomic features identified as being significantly associated with a clinical endpoint may be included in an initial or final radiomic signature. Alternatively, only a subset of radiomic features significantly associated with a clinical endpoint may be included in an initial or final signature, for example, at least two. Alternatively, at least three of the radiomic features significantly associated with a clinical endpoint may be included in an initial or final signature. Alternatively, at least four of the radiomic features significantly associated with a clinical endpoint may be included in an initial or final signature. Alternatively, at least five of the radiomic features significantly associated with a clinical endpoint may be included in an initial or final signature. Alternatively, at least six of the radiomic features significantly associated with a clinical endpoint may be included in an initial or final signature. Alternatively, at least seven of the radiomic features significantly associated with a clinical endpoint may be included in an initial or final signature.

[0204] Likewise, labels are preferably refined using the data used for the training cohort, and then validated externally using the data used for the training cohort.

[0205] Elements of clustering methods or non-clustering methods may be combined. For example, a clustering (or collinear elimination) method may involve the application of a Bonferroni correction.

[0206] For example, multiple radiomic features can be input into a machine learning algorithm without performing some (or any) of the aforementioned steps described above. However, typically, at least the step of eliminating unstable radiomic features is performed before inputting the features into the machine learning algorithm.

[0207] Typically, in the above methods (both clustered and non-clustered), bivariate associations between radiomic features can be assessed by the non-parametric Spearman's rho (ρ) coefficient, while observer variability can be assessed by the intraclass correlation coefficient (ICC) within a given number of scans (e.g., at least two scans, e.g., ten scans).

[0208] Radiomics signature

[0209] The PVR radiomics signature of the present invention is calculated based on the measured values ​​of the radiomics features obtained from the medical imaging data. Specifically, the PVR radiomics signature is preferably calculated based on at least two radiomics features.

[0210] To improve the prognostic and diagnostic value of the signature, as described above, the signature is preferably calculated based on at least two radiomic features selected from different "clusters" of collinear 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 PVR.

[0211] Nine clusters have been identified using the "clustering" method, corresponding to the nine "original" or "partner" features in the reduced multiple radiomic features that survived the collinear elimination process in the study described in the "Example" section below. Table 1 identifies the members of the nine clusters.

[0212] Table 1 also identifies the "original features", which are the features in each cluster that survived the elimination of collinear features in the "clustering" method described above. The collinearity of other features in each cluster with the original features is represented by the |ρ| coefficient (where ρ is the nonparametric Spearman rho (ρ) coefficient).

[0213] To construct the clusters of Table 1, 53 stable volume- and orientation-independent radiomic features that were associated with clinical endpoints above a threshold (p<0.05; where p is a probability value or asymptotic significance, i.e., α=0.05) in both the training and validation cohorts were clustered based on their strength of association with the original features. If a radiomic feature was found to be collinear with two or more original features (|rho|≥0.75), it was assigned to the cluster of the original feature with which it was most closely associated. Interestingly, and perhaps surprisingly, FAI PVAT is not associated with any of the original features at the level of |rho|≥0.75 and is therefore not included in any of the clusters. This shows that PVAT The present invention provides a separate and distinct prognostic and diagnostic tool compared to FAI. PVAT Complementary.

[0214] Table 1: Radiomic features of the “standard” cluster were selected from the 53 identified features significantly associated with MACE

[0215]

[0216]

[0217] As mentioned previously, clusters can be expanded to include stable members of the original multiple radiomic features that are collinear with the “original” radiomic features of the cluster, regardless of whether the collinear radiomic features themselves are independently significantly associated with the clinical endpoint.

[0218] Table 2 presents the membership of the nine "expanded" clusters. These expanded clusters are equivalent to the original clusters presented in Table 1, but are expanded to include any radiomic features from the 719 stable radiomic features (ICC>0.9) that are also collinear with one of the original features from each cluster (|rho|≥0.75). Likewise, if a radiomic feature is found to be collinear with two or more original features (|rho|≥0.75), the radiomic feature is assigned to the cluster of the original feature with which it is found to be most closely associated.

[0219] Table 2: Radiomic features of the “expanded” cluster were selected from 719 identified stable features

[0220]

[0221]

[0222]

[0223]

[0224] The radiomic signature can be constructed from any of the radiomic signatures included in Table 1 or Table 2 (i.e., standard clustering or extended clustering), provided that the radiomic signature includes at least two radiomic signatures each selected from different clusters. For example, the radiomic signature may include at least three radiomic signatures from different clusters. For example, the radiomic signature may include at least four radiomic signatures from different clusters. For example, the radiomic signature may include at least five radiomic signatures from different clusters. For example, the radiomic signature may include at least six radiomic signatures from different clusters. For example, the radiomic signature may include at least seven radiomic signatures from different clusters. For example, the radiomic signature may include at least eight radiomic signatures from different clusters. For example, the radiomic signature may include at least nine radiomic signatures from different clusters.

[0225] Each of the at least two (or more) radiomic features selected from different clusters can be selected to be correlated with the "original" features of the cluster to which it belongs at least at a degree of |rho|=0.8. For example, each of the at least two radiomic features from different clusters can be correlated with the "original" features of the cluster to which it belongs at least at a degree of |rho|=0.85. For example, each of the at least two radiomic features from different clusters can be correlated with the "original" features of the cluster to which it belongs at least at a degree of |rho|=0.9. For example, each of the at least two radiomic features from different clusters can be correlated with the "original" features of the cluster to which it belongs at least at a degree of |rho|=0.95.

[0226] In addition to calculating a radiomic signature based on at least two radiomic features from different clusters, the radiomic signature can also be calculated based on additional radiomic features. For example, a radiomic signature can include more than one radiomic feature from any given cluster, or can include radiomic features that are not included in any of the clusters.

[0227] Alternatively, the radiomic signature may be calculated based on at least two of the radiomic signatures that have been found to be independently associated with the clinical endpoint using a "non-clustering" approach. Thus, the radiomic signature may alternatively be calculated based on at least two radiomic signatures selected from Table 3.

[0228] Table 3: Radiomic features identified as significantly associated with MACE according to the “non-clustering” approach

[0229]

[0230]

[0231] The radiomic signature may be calculated based on at least two radiomic features from Table 3. For example, the radiomic signature may be calculated based on at least three radiomic features from Table 3. For example, the radiomic signature may be calculated based on at least four radiomic features from Table 3. For example, the radiomic signature may be calculated based on at least five radiomic features from Table 3. For example, the radiomic signature may be calculated based on at least six radiomic features from Table 3. For example, the radiomic signature may be calculated based on at least seven radiomic features from Table 3.

[0232] At least two (or more) radiomic features can be selected from those radiomic features in Table 3 with a Bonferroni-adjusted P value <0.04. For example, a radiomic feature with a Bonferroni-adjusted P value <0.03. For example, a radiomic feature with a Bonferroni-adjusted P value <0.02. For example, a radiomic feature with a Bonferroni-adjusted P value <0.01. For example, a radiomic feature with a Bonferroni-adjusted P value <0.005. For example, a radiomic feature with a Bonferroni-adjusted P value <0.002.

[0233] Signatures can be constructed from the radiomic features listed in Table 3 , except for the mean.

[0234] Likewise, in addition to calculating a radiomic signature based on at least two radiomic features selected from the radiomic features presented in Table 3, the radiomic signature may also be calculated based on additional radiomic features. For example, a radiomic signature may include radiomic features not included in Table 3.

[0235] Each of the radiomic signatures of the present invention provides a direct way to characterize PVR using medical imaging data. Because each of the radiomic signatures 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.

[0236] system

[0237] 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.

[0238] Specifically, methods for deriving radiomic signatures can be performed on such systems, and thus such systems are provided according to the present invention. For example, the system can be configured to receive and optionally store a data set comprising values ​​of multiple radiomic features of PVR obtained from medical imaging data for each of a plurality of individuals, and information about what occurred during a subsequent time period after collecting medical imaging data indicating a clinical endpoint of cardiovascular risk for each of a plurality of individuals. The system can be configured to construct (e.g., derive and validate) a radiomic signature using such a data set according to the methods of the present invention.

[0239] Alternatively, the system may be configured to perform a method of characterizing a PVR. Specifically, the present invention provides a system for characterizing a PVR using medical imaging data of a subject. The system may be configured to calculate a value of a radiomics signature of the PVR using the medical imaging data. The radiomics signature may be calculated based on measured values ​​of at least two radiomics features of the PVR, and the measured values ​​of at least two radiomics features may be calculated based on the medical imaging data.

[0240] 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.

[0241] Definition of radiomic features

[0242] 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.

[0243] a. First-order statistics

[0244] 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. In this way, the statistics describe the quantitative and qualitative characteristics of the entire ROI (PVR). A total of 19 features were calculated for each of the eight wavelet transforms and the original CT image, as follows:

[0245] make:

[0246] 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)

[0247] 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.

[0248] p(i) is the normalized first-order histogram and is equal to

[0249] 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 PVR (-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.

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

[0251] Table R1: First-order radiomic features for PVR characterization

[0252]

[0253]

[0254] b. Shape-related statistics

[0255] 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:

[0256] make:

[0257] V is the volume of ROI, in mm 3For units

[0258] A is the surface area of ​​ROI, in mm 2 Unit

[0259] Table R2: Shape-related radiomic features for PVR characterization

[0260]

[0261]

[0262] c. Gray Level Co-occurrence Matrix (GLCM)

[0263] 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.

[0264] make:

[0265] ∈ is an arbitrarily small positive number (e.g., about 2.2×10 -16 )

[0266] P(i,j) is the co-occurrence matrix of any δ and θ

[0267] p(i,j) is the normalized co-occurrence matrix and is equal to

[0268] Ng is the number of discrete intensity levels in the image

[0269] is the marginal row probability

[0270] is the marginal column probability

[0271] μ x For p x The average gray intensity of

[0272] μ y forp y The average gray intensity of

[0273] σ x Yes x Standard Deviation

[0274] σ y Yes y Standard Deviation

[0275] Where i+j=k, and k=2, 3, ..., 2N g

[0276] where |ij|=k, and k=0, 1, ..., N g -1

[0277] Yes x Entropy

[0278] Yes y Entropy

[0279]

[0280]

[0281] For distance weighting, the GLCM matrix is ​​weighted by the weighting factor W, and then summed and normalized. By W = e - ‖d‖ 2 A weighting factor W is calculated for the distance between neighboring voxels, where d is the distance of the associated angle.

[0282] Table R3: Gray Level Co-occurrence Matrix (GLCM) Statistics for PVR Characterization

[0283]

[0284]

[0285]

[0286] d. Grayscale Size Zone Matrix (GLSZM)

[0287] 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.

[0288] make:

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

[0290] N s is the number of discrete region sizes in the image

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

[0292] N z is the number of regions in the ROI, which is equal to and

[0293] 1≤N z ≤N p

[0294] P(i,j) is the size region matrix

[0295] p(i,j) is the normalized size region matrix, defined as

[0296] ∈ is an arbitrarily small positive number (e.g., about 2.2×10 -16 ).

[0297] Table R4: Grayscale size zone matrix (GLSZM) statistics for PVR characterization

[0298]

[0299]

[0300]

[0301] e. Grayscale run length matrix (GLRLM)

[0302] 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.

[0303] make:

[0304] N g is the number of discrete intensity values ​​in the image

[0305] N r is the number of discrete run lengths in the image

[0306] N p is the number of voxels in the image

[0307] N z (θ) is the number of runs along angle θ in the image, equal to And 1≤N z (θ)≤N p

[0308] P(i,j|θ) is the run length matrix of any direction θ

[0309] p(i,j|θ) is the normalized run length matrix, defined as

[0310] ∈ is an arbitrarily small positive number (e.g., about 2.2×10 -16 ).

[0311] 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".

[0312] Table R5: Gray-level run length matrix (GLRLM) statistics for PVR characterization

[0313]

[0314]

[0315]

[0316] f. Neighboring Gray Tone Difference Matrix (NGTDM) features

[0317] 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:

[0318]

[0319] 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

[0320] Here, W is also in X gl The number of voxels in the neighborhood of .

[0321] make:

[0322] n i is X with gray level i gl The number of voxels in

[0323] 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.

[0324] p i is the gray level probability and is equal to n i / N v

[0325]

[0326] N g is the number of discrete gray levels

[0327] N g,p For p iThe number of gray levels when ≠0

[0328] Table R6: Neighboring Gray Tone Difference Matrix (NGTDM) for PVR Characterization

[0329]

[0330]

[0331] g. Gray Level Dependency Matrix (GLDM)

[0332] 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 that has j related voxels in its neighborhood occurs in the image.

[0333] N g is the number of discrete intensity values ​​in the image

[0334] N d is the number of discrete dependency sizes in the image

[0335] N z is the number of dependent regions in the image, which is equal to

[0336] P(i,j) is the dependency matrix

[0337] p(i,j) is the normalized dependency matrix, defined as

[0338] Table R7: Gray Level Dependence Matrix (GLDM) Statistics for PVR Characterization

[0339]

[0340]

[0341]

[0342] Examples

[0343] method

[0344] A two-group study was conducted to explore the diagnostic and prognostic value of radiomic phenotyping of coronary PVR on CCTA. Figure 1 and Tables 4A, 4B, and 5 present the study flow chart and baseline characteristics of study groups 1 and 2. The study protocol was approved by all local institutional review boards.

[0345] Group 1

[0346] This was a case-control study of patients with adverse events versus those without adverse events within five years after a clinically indicated assessment by CCTA. Eligible cases were retrieved from prospectively collected data of two independent cohorts of patients undergoing clinically indicated CCTA. Of a total of 4239 individual scans reviewed (2246 in cohort 1 and 1993 in cohort 2), 3912 were of excellent quality and included in additional analyses. Cases were identified based on the primary composite endpoint of major adverse cardiovascular events (MACE), defined as the composite of all-cause mortality and non-fatal myocardial infarction (MI) within five years after CCTA, while controls were identified as patients with event-free follow-up of at least five years after CCTA. After reviewing scan quality and retrieving relevant demographic characteristics, 1:1 matching was performed for age, sex, obesity status, cohort, and technical parameters (tube voltage and CT scanner used) related to CCTA acquisition using an automated algorithm. Patients were also matched for other cardiovascular risk factors, including hypertension, dyslipidemia, diabetes, and smoking, when possible. Subgroups of patients with cardiac-specific MACE (cMACE; cardiac mortality and nonfatal MI) and their matched controls were also examined separately to increase the sensitivity of cardiac-specific high-risk PVR signatures.

[0347] Table 4A: Demographic characteristics of study group 1

[0348]

[0349]

[0350] Values ​​are presented as percentage (%) or mean ± SD.

[0351] * Presented as a valid percentage.

[0352] ACEi: angiotensin-converting enzyme inhibitor; ARB: angiotensin II-receptor blocker; CAD: coronary artery disease; (c) MACE: (cardiac-specific) major adverse cardiovascular event.

[0353] Table 4B: Demographic characteristics of study group 1

[0354]

[0355]

[0356] Values ​​are presented as percentage (%) or mean ± SD.

[0357] * Presented as a valid percentage.

[0358] ACEi: angiotensin-converting enzyme inhibitor; ARB: angiotensin II-receptor blocker; CAD: coronary artery disease; (c) MACE: (cardiac-specific) major adverse cardiovascular event.

[0359] Group 2

[0360] This was a prospective study (Ox-IMPACT study, with ethical approval from the Oxford C Research Ethics Committee South Central; REC reference 17 / SC / 0058) recruiting 22 patients with acute myocardial infarction (n=22 unstable lesions) who were invited to undergo serial CCTA scans 96 hours and six months after admission. A control group of 32 patients with known stable CAD (n=39 stable lesions) and with prior percutaneous coronary intervention (PCI) at least three months before the CCTA scan were also included in this group. Radiomic phenotyping of coronary PVR was performed around both stable and unstable lesions to identify radiomic signatures of PVR associated with plaque instability and inflammation.

[0361] Table 5: Demographic characteristics of study group 2

[0362]

[0363] Radiomic features included in the study

[0364] A total of 843 PVR radiomic features were measured, as summarized in Table 6 .

[0365] Table 6: Subdivision of PVR radiomic features

[0366]

[0367] Data collection, definitions, and outcome assessment

[0368] In group 1, outcome data were compiled by searching medical records and querying local / national databases by local investigators who were not involved in the subsequent image / data analysis. Appropriate institutional review board approval was obtained with a waiver of individual informed consent. Clinical data and demographic characteristics were pre-recorded in the electronic medical record at the time of the initial clinical encounter and were manually extracted for the current study. Hypertension was defined based on the presence of a documented diagnosis or treatment of antihypertensive regimen according to relevant clinical guidelines (James PA, Oparil S, Carter BL, et al., 2014. Evidence-based guidelines for the management of hypertension in adults: a report of the eighth Joint National Committee (JNC 8) appointed panelists. JAMA 2014;311(5):507-20). Similar criteria were applied to the definition of hypercholesterolemia and diabetes (American Diabetes A. Diagnosis and classification of diabetes mellitus. Diabetes Care 2014;37 Suppl 1:S81-90; Stone NJ, Robinson JG, Lichtenstein AH, et al., 2013 ACC / AHA guideline on cholesterol therapy to reduce the risk of atherosclerotic cardiovascular disease in adults: a report of the American College of Cardiology / American Heart Association Task Force on Practice Guidelines. J Am Coll Cardiol 2014;63(25Pt B):2889-934). The exact cause of death was determined by investigators at each site through medical record review, death certificate review, and / or telephone follow-up and / or verification with family members. Cardiac and noncardiac mortality were defined according to the recommendations of the ACC / AHA and the Academic Research Consortium as previously described. More specifically, cardiac mortality was defined as any death due to adjacent cardiac causes (e.g., myocardial infarction, low-output heart failure, fatal arrhythmia). Deaths that met criteria for sudden cardiac death were also included in this group. Any death not covered by the previous definition, such as death due to malignancy, accident, infection, sepsis, renal failure, suicide, or other non-cardiovascular causes (e.g., stroke or pulmonary embolism) was classified as non-cardiac. In cases where it was not possible to collect definitive information about the cause of death, the death was classified as "cause of death unknown" by the local site investigator. Non-fatal myocardial infarction events (ST-segment elevation or non-ST-segment elevation) during the follow-up period can also be retrieved through searches of the electronic health records.

[0369] Coronary CT Angiography (CCTA) Acquisition Protocol

[0370] Queue 1: The majority of CCTA scans (87.1%) were performed using a 256-slice Brilliance iCT scanner (Philips Medical Systems, Best, The Netherlands), with the remainder being performed using a 2×128-slice Definition Flash scanner (Siemens Healthcare, Erlangen, Germany) (10.8%) and a 2×192-slice Somatom Force CT scanner (Siemens Healthcare, Forchheim, Germany) (2.1%). For patients with a heart rate >60 beats / min, 5 mg of metoprolol (titrated in 5 mg increments to a maximum of 30 mg) or diltiazem (titrated in 5 mg increments to a maximum of 20 mg) were administered intravenously if the patient's heart rate remained above 60 beats per minute once positioned on the CT table. Patients also received 0.3 mg nitroglycerin sublingually immediately before CCTA and administration of iodinated contrast medium (Omnipaque 350, General Electric, Milwaukee, USA) at a flow rate of 5-6 ml / sec.

[0371] Queue 2: CCTA scans were performed on a 2×64 slice scanner (Definition Flash, Siemens Medical Devices, Forchheim, Germany) (79.2%), with the remainder performed using a 64 slice (Siemens Sensation 64, Siemens Medical Devices, Forchheim, Germany) (18.1%) or 2×128 slice scanner (2.7%) (Somatom Definition Flash, Siemens Medical Devices, Forchheim, Germany). If the patient's heart rate remained above 60 beats per minute once positioned on the CT table, an oral medication containing 100 mg of atenolol was administered one hour before CT if the heart rate was >60 beats per minute with an additional 5 mg dose of metoprolol intravenously up to a maximum dose of 30 mg. Patients also received 0.8 mg of nitroglycerin sublingually immediately before CCTA and administration of iodinated contrast agent (Omnipaque 350, Schering AG, Berlin, Germany) at a flow rate of 5-6 ml / sec.

[0372] Study Group 2: CCTA was performed on participants in study group 2 using a 64-slice scanner (General Electric, LightSpeed ​​Ultra, General Electric, Milwaukee, WI, USA). Heart rate was optimized using intravenous beta-blockers, and glyceryl trinitrate (800 μg) was also administered sublingually to achieve maximum coronary vasodilation. CCTA (tube energy of 120 kVp, axial slice thickness of 0.625 mm, rotation time of 0.35 seconds, and detector coverage of 40 mm) was performed after intravenous injection of 95 ml of iodine-based contrast medium (Niopam 370, BRACCO UK Ltd) at a flow rate of 6 ml / sec. Prospective image acquisition was used at 75% of the cardiac cycle by ECG-gating (if necessary, the right coronary artery was optimally imaged using a fill time of 100 milliseconds).

[0373] CCTA scan processing and analysis

[0374] All images were first anonymized locally and subsequently transferred to a core laboratory (Academic Cardiovascular CT Unit, University of Oxford, United Kingdom) for analysis by investigators blinded to demographic characteristics and outcomes on a dedicated workstation (Aquarius V.4.4.13, TeraRecon Inc., Foster City, CA, USA. Initially, all scans were reviewed based on their quality and the presence of artifacts that precluded reliable qualitative and quantitative assessments. Scans of low quality were reviewed by at least two independent investigators and then excluded from subsequent analyses. Mild, moderate, and severe coronary artery stenosis were defined as luminal narrowing of 25–49%, 50–69%, and ≥70%, respectively (as previously described in Cury RC, Abbara S, Achenbach S, et al., CAD-RADS(TM) Coronary Artery Disease-Reporting and Data System, an expert consensus document of the Society of Cardiovascular Computed Tomography (SCCT), the American College of Radiology (ACR), and the North American Society for Cardiovascular Imaging (NASCI), supported by the American College of Cardiology, J Cardiovasc Comput Tomogr 2016;10(4):269–81). Obstructive coronary artery disease (CAD) was defined as ≥1 coronary artery lesion causing ≥50% luminal stenosis, and the extent of CAD was assessed by the Duke Prognostic CAD Index (e.g., as defined in Min JK, Shaw LJ, Devereux RB, et al., Prognostic value of multidetector coronary computed tomographic angiography for prediction of all-cause mortality. J Am Coll Cardiol 2007;50(12):1161-70).High-risk plaque features were defined as the presence of at least one of the following features on CCTA: a) speckled calcification, b) low-attenuation plaque, c) positive remodeling, and d) napkin ring sign (as previously defined in Puchner SB, Liu T, Mayrhofer T, et al., High-risk plaque detected on coronary CT angiography predicts acute coronary syndromes independent of significant stenosis in acute chest pain: results from the ROMICAT-II trial, J Am Coll Cardiol 2014;64(7):684-92). Epicardial (visceral) obesity was assessed by measuring total epicardial adipose tissue (EAT) volume by semiautomatically tracing the pericardial contour from the level of the pulmonary artery bifurcation to the most caudal apex of the heart. FAI. PVAT As previously defined in Antonopoulos AS, Sanna F, Sabharwal N, et al., Detecting human coronary inflammation by imaging perivascular fat, Sci Transl Med 2017;9(398).

[0375] Coronary PVR radiomics feature extraction

[0376] Calculation of radiomic features of coronary PVR was performed in all selected CCTA scans using 3D Slicer (v.4.9.0-2017-12-18r26813, available from http: / / www.slicer.org). To avoid collinearity issues between different coronary vessels, the analysis of Group 1 was limited to the proximal right coronary artery and the middle right coronary artery (RCA) (segment 1 and segment 2 according to the anatomical classification of the American Heart Association). Coronary PVR was defined as all voxels located within a radial distance from the outer vessel wall equal to the diameter of the adjacent vessel within the range of -190 to -30 Hounsfield units (HU). Segmentation of coronary PVR was performed by placing a three-dimensional sphere with a diameter equal to three times the diameter of the coronary vessel on continuous slices along the centerline of the vessel. Subsequently, the segmented PVR was extracted and used to calculate radiomic features using the SlicerRadiomics extension of 3D Slicer, which incorporates the Pyradiomics library into 3D Slicer. The raw HU values ​​of the segmented PVR were used to calculate shape-related radiomic features and first-order radiomic features. To calculate texture features (gray-level co-occurrence matrix [GLCM], gray-level dependency matrix [GLDM], gray-level run length matrix [GLRLM], gray-level size region matrix [GLSZM], and neighboring gray-level tone difference matrix [NGTDM], Figure 1A and Tables R1-R7), discretizing PVR voxels into 16 equally wide bins (ten HU width) to reduce noise while allowing sufficient resolution to detect biologically significant spatial changes in PVR decay. To implement symmetric rotation-invariant results, texture statistics (GLCM, etc.) can be calculated in all four directions and then averaged (e.g., as previously described in Kolossvary M, Kellermayer M, Merkely B, Maurovich-Horvat P., Cardiac Computed Tomography Radiomics: A Comprehensive Review on Radiomic Techniques, Journal of thoracic imaging 2018; 33(1): 26-34).

[0377] Wavelet Transform: First-order statistics and texture-based statistics were also calculated for the 3D wavelet transform of the raw image data, resulting in eight additional radiomic feature sets ( Figure 1A ).

[0378] Statistical analysis

[0379] In group 1, case-control matching was performed using an automated algorithm provided by the ccmatch command in Stata. In the final study population, clinical demographic characteristics are presented as mean ± standard deviation for continuous variables and percentages for categorical variables. Continuous variables between the two groups were compared by Student's t-test, while categorical variables were compared by Pearson's Chi-square test.

[0380] Principal components and unsupervised clustering: In both study groups, all 843 calculated PVR radiomics features were included in principal component analysis to identify principal components describing the majority of phenotypic variation in the study population. The observations were then hierarchically clustered (using Ward's method and Minkowski distance) using the three first components for group 1 (PC1, PC2, PC3) and group 2 (PC1', PC2', and PC3'). The frequencies of MACE (group 1) and unstable plaques (group 2) were compared between the different clusters by chi-square test.

[0381] Feature selection and improved discriminability: First, the discriminatory value of all radiomic features for 5-year MACE was tested in receiver operating characteristic (ROC) curve analysis. This analysis was performed on pooled data from both cohorts. To correct for multiple comparisons, a genome-wide association study (GWAS)-based approach previously used in the field of radiomic image analysis was followed, with a Bonferroni correction for the number of principal components accounting for 99.5% of the variability in the present study samples (see, e.g., Kolossvary M, Karady J, Szilveszter B, et al., Radiomic signatures outperform conventional quantitative computed tomography metrics for identifying coronary artery plaques with napkin ring sign, Circ Cardiovasc Imaging 2017;10(12) and Johnson RC, Nelson GW, Troyer JL, et al., Accounting for multiple comparisons in a genome-wide association study (GWAS), BMC Genomics 2010;11:724). Bivariate associations between radiomic features could be assessed by the nonparametric Spearman's rho (ρ) coefficient, while observer variability could be assessed by the intraclass correlation coefficient (ICC) across the ten scans.

[0382] To construct a radiomic signature for high-risk PVR, a multi-step approach was used. Z-score transformation was applied to all features, and unstable radiomic features with low ICC in repeated analyses (<0.9) were excluded. In order to minimize false positive results caused by cohort-specific changes, features significantly associated with MACE in both cohorts were selected (at the α=0.05 level). Next, collinearity was reduced by stepwise removal of pairwise comparisons using the appropriate functions of the caret package on R. Using the remaining nine radiomic features, a machine learning method using elastic network regression and leave-one-out internal cross-validation was applied in cohort 1. The optimal penalty coefficient (λ, λ) was selected by cross-validation, and α was preferably set to α=1. Then, the best performing models for both MACE and cMACE were externally validated in cohort 2, and the discrimination was evaluated by calculating the area under the curve (AUC). Then, the best variables of the best performing models were combined into a unified score / label by multiplying each model by the corresponding unadjusted β coefficient of the model and calculating the sum. The unified labels were then added to a logistic regression model consisting of the following four blocks: i) age, sex, hypertension, dyslipidemia, smoking, and diabetes (model 1); ii) model 1 + CT-derived measures containing the Duke Prognostic CAD Index, the presence of high-risk plaque features, EAT volume, and the presence of coronary artery calcium (model 2); iii) model 2 + FAI PVAT (Model 3); and iv) Model 3 + PVR texture signature. The prognostic value of the nested models for MACE and cMACE was compared by their corresponding C statistics (area under the curve, AUC). The FAI was graphically presented using bidirectional contour plots derived from previous multiple logistic regression models. PVAT Interactions with PVR texture labels. Selected radiomic features between unstable and stable plaques in group 2 were compared using the nonparametric Mann-Whitney test. Statistical analyses were performed in the R environment (packages: caret, hclust) and Statav14.0 (Stata Collaboration, College Station, Texas). All tests were two-sided and α was set to 0.05 unless otherwise stated.

[0383] result

[0384] Demographic characteristics of the study population

[0385] Of the 3912 individual scans reviewed in Group 1, a total of 386 scans were included in the study, corresponding to 193 5-year MACE patients and 193 matched controls. In the subgroup type analysis, a selected subgroup of 98 cardiac-specific MACE (cMACE) patients and their matched controls were also identified and analyzed. Tables 4A and 4B summarize the demographic and baseline characteristics of the Study Group 1 population. Cases and controls did not differ significantly in their baseline demographic characteristics, however, as expected, cases (MACE or cMACE) were more likely to have coronary artery disease (CAD) compared to their controls, as assessed by the degree of coronary artery stenosis and the Duke Prognostic CAD Index. Study Group 2 included 22 patients with unstable lesions and 32 age- and sex-matched controls with 39 stable lesions (Table 5).

[0386] PVR radiomics signature: component analysis and association with adverse events

[0387] A total of 103 radiomic features were calculated from the original images, corresponding to 15 shape-related features in addition to 18 first-order statistics, 23 GLCM, 14 GLDM, 16 GLRLM, 16 GLSZ, and 5 NGTDM individual features ( Table 6 , Figure 4 All features except shape-related features were also calculated for each of the eight wavelet transforms of the segmented regions, yielding a total of 843 radiomic features. Principal component analysis of all radiomic features revealed that three principal components accounted for 61.98% of the variation observed in both cohorts of Group 1 ( Figure 5A ), and 92 components accounted for 99.5% of the variation in the study population ( Figure 5B A total of six components among the 42 components with eigenvalues ​​> 1 were significant predictors of MACE ( Figure 5C ), whereas in a multivariate regression model including all five components adjusted for age, sex, cardiovascular disease risk factors, and CCTA-derived CAD severity index, four components remained independently associated with MACE ( Figure 5D These components were associated with baseline clinical demographics and characteristic differences ( Figure 5E ), likely reflecting PVR signatures that describe different biological phenotypes. Notably, inclusion of the four PVR components in a model that included age, sex, cardiovascular disease risk factors, and a CCTA-derived risk signature significantly improved discrimination for 5-year MACE ( Fig. 5F ), thus indicating that radiomic phenotyping of PVR has a strong and incremental prognostic value.

[0388] Unsupervised clustering based on coronary PVR phenotyping

[0389] Unsupervised (hierarchical) clustering of the pooled cohort 1 study population using the first three principal components of coronary PVR radiomics (PC1, PC2, and PC3) identified three distinct clusters with significantly different 5-year MACE risks (46.5% vs. 45.8% vs. 65.8% MACE, P=0.009). Similarly, in cohort 2, hierarchical clustering of the identified coronary lesions identified two distinct clusters with significantly different prevalence of unstable plaques (58.8% vs. 25%, P=0.01). These findings suggest that there are distinct radiomic signatures in PVR that are associated with increased cardiovascular disease and the local presence of coronary inflammation and unstable lesions.

[0390] Identification of specific high-risk PVR radiomic features

[0391] Principal component analysis and unsupervised clustering proved the concept that radiomic phenotyping of PVR can be associated with both the local presence of coronary inflammation / disease and poorer outcome. However, it failed to identify specific high-risk radiomic signatures that could be reproducibly measured in independent cohorts. Principal components are specific to the dataset from which they were derived and are not easily applicable to independent datasets and cohorts. Therefore, additional work was performed to identify FAIs that could complement the established PVAT Specific features of the diagnostic and prognostic information of the markers. In the ROC analysis of the primary endpoint for distinguishing MACE, using the pooled data of the two cohorts, a total of 198 features were found to be significant at the 0.05 level. In order to correct for multiple comparisons and reduce the false discovery rate (FDR), a Bonferroni correction was applied based on principal component analysis (new significance cutoff = 0.05 / 92 = 0.00054347826). After this correction, only 46 radiomics features remained significant discriminants of MACE, as shown in the Manhattan plot ( Figure 6 ) and summarized in Table 3. Of these 46 radiomic features, 28 were first-order statistics, including 10 GLRLM, 5 GLCM, 1 GLDM, 1 GLSZ, and 1 NGTDM, while half were derived from the original images and the other half were derived from wavelet transform (17 LLL, 2 HHH, 3 HHL, and 1 LHL).

[0392] Constructing a high-risk PVR radiomics signature

[0393] Although FAI PVATIndividual radiomic features such as were associated with increased cardiovascular risk, but it was previously unclear whether a combination of different radiomic features (“signatures”) could provide a more powerful way to characterize the adverse profile of coronary PVR. To explore this hypothesis and develop a prognostic and reproducible radiomic signature, a rigorous stepwise approach was applied ( Figure 2 ). First, of the 843 coronary PVR radiomic signatures calculated, 143 (ICC < 0.9, two measurements) were excluded after stability analysis ( Figure 7 ). Next, 60 volume- and orientation-independent radiomic features that were significantly associated with MACE at the a=0.05 level in both cohorts were selected to reduce possible false-positive findings due to cohort-specific variation. Collinearity was subsequently reduced by stepwise elimination of pairwise correlations at the level of |rho|≥0.75 (Spearman rho). Using the remaining nine radiomic features, a machine learning approach using elastic net / lasso regression and leave-one-out internal cross-validation was applied in cohort 1. The best performing model for predicting MACE showed an average discrimination power ( Figure 2B ), and a similar model for predicting cardiac-specific MACE ( Figure 2C ) showed good discriminative value in both cohort 1 (derivation cohort) and cohort 2 (validation cohort). The coefficients derived from the model were then used to define the perivascular texture index (PTI) using the first six components of this model, as shown in Table 7. Thus, Table 7 describes a specific example of the radiomics signature of the present invention.

[0394] Table 7: Definition of high risk PVR texture labels

[0395]

[0396] In Table 7, by z The z-score β was converted to an unstandardized β (b i ). The constant 90 was added post hoc to ensure a positive value based on the range of values ​​observed in all analyzed cohorts.

[0397] The incremental value of PVR radiomic phenotyping beyond current state-of-the-art techniques

[0398] To evaluate FAI PVATTo demonstrate the incremental value of radiomic phenotyping beyond current risk biomarkers used in CCTA-based risk stratification, a set of nested models were constructed, as shown in Figures 3A and 3B. Model 1, consisting of demographic characteristics and conventional risk factors, showed poor discrimination for both MACE and cMACE, in part due to the matching process used to select the control group. Addition of a CCTA-derived risk signature (comprising the Duke Prognostic CAD Index, the presence of coronary artery calcium, the presence of high-risk plaque features, and EAT (epicardial adipose tissue) volume) significantly improved discrimination for both MACE and cMACE, whereas inclusion of FAI PVAT Despite this significant improvement, the addition of PTI (as defined in Table 7) was associated with a further significant improvement in the model's discrimination for both endpoints (Figures 3A and 3B), suggesting that PVR texture phenotyping has both independent and incremental value in risk prediction. Figure 3C The FAI is graphically presented in PVAT The interaction between FAI and PTI was used to predict MACE / cMACE, which showed that for a given FAI PVAT values, there is an increased risk when PTI levels are higher, and vice versa.

[0399] Validation of the present invention's alternative radiomic signature (PTI)

[0400] The data presented in Figures 3A and 3B demonstrate that the radiomic signature (PTI) defined in Table 7 provides significantly improved discriminatory value of the model for both endpoints (MACE and cMACE). To validate the usefulness of the alternative radiomic signatures of the present invention comprising different selections of radiomic features, a series of models comprising several different radiomic signatures were tested against prior art models.

[0401] In Table 8, the model performance in 196 patients (98 patients with cardiac MACE and 98 matched controls) is improved. Each step corresponds to the inclusion of a selected radiomics feature and previously clustered radiomics features on top of the prior art model. The prior art model included age, sex, hypertension, dyslipidemia, smoking and diabetes, Duke Prognostic Coronary Artery Disease Index, high-risk plaque features, epicardial adipose tissue volume, and the presence of coronary calcium.

[0402] In each of Examples 1-4, the prior art model is gradually supplemented by a radiomic signature containing increasingly more radiomic features from different clusters. First, the prior art model is supplemented by a radiomic signature calculated based on a radiomic feature selected from Cluster 1 (the first row of Tables 8A and 8B: "+ Cluster 1"). Next, the prior art model is supplemented by a radiomic signature calculated based on two radiomic features selected from Clusters 1 and 2 (the first row of Tables 8A and 8B: "+ Cluster 2"). Therefore, each row of Tables 8A and 8B corresponds to a radiomic signature containing a selected radiomic feature and a radiomic feature of a previous cluster on top of the prior art model. Nagelkerke's pseudo-R2 provides a measure of the discriminatory power of the model for cMACE.

[0403] It is clear from Tables 8A and 8B that the signatures of the present invention calculated based on different selections of radiomic features from the identified clusters all provide improved predictions for cardiac-specific MACE. Therefore, the data presented in Table 8 show that no matter which features are selected from each of the identified clusters, or exactly how many features are selected, the radiomic signatures of the present invention (PTI) provide improved predictions for cardiovascular risk compared to previously used models.

[0404] Table 8A: Improved prediction of cardiac-specific MACE by different selections of radiomic features from the identified clusters

[0405]

[0406]

[0407] Table 8B: Improved prediction of cardiac-specific MACE by different selections of radiomic features from the identified clusters

[0408]

[0409] Radiomic phenotyping of PVR to detect unstable lesions

[0410] When calculated around predefined coronary lesions in study group 2, all radiomic features except the identified radiomic features used to define PTI were significantly altered in the presence of unstable culprit lesions (scanned within 96 hours of ACS onset) compared with stable treated lesions (scanned within >3 months after PCI) (Figure 8A-G). More importantly, PTI (defined using the formula in Table 7) outperformed all six individual radiomic features in distinguishing unstable coronary lesions from stable coronary lesions (AUC: 0.76; 95% confidence interval; 0.64-0.88) (Figure 8G). These findings suggest that the combination of PTI and FAI can improve the quality of life of patients with ACS. PVAT Perivascular fat phenotype analysis is valuable for both future risk prediction and detection of unstable coronary plaques and reveals a close association between perivascular fat phenotype and vascular inflammation, plaque instability, and adverse clinical outcomes.

[0411] Summary of findings

[0412] Using machine learning methods, the inventors discovered and validated that a coronary PVR radiomic signature adds incremental value beyond traditional risk factors and established CCTA risk classification tools in predicting future adverse cardiovascular events and assessing cardiovascular health and risk, and further detects the presence of localized plaque inflammation and unstable coronary artery lesions. The inventors have demonstrated that a PVR radiomic signature based on two or more radiomic features of PVR provides a tool for predicting future adverse events in patients, for diagnosing coronary artery disease or coronary heart disease, and for identifying unstable coronary artery lesions.

[0413] The PVR signature of the present invention describes a high-risk PVR phenotype, links it to future event risk, and provides incremental prognostic information beyond current CCTA-based tools. The signature of the present invention is also capable of distinguishing unstable coronary lesions from stable coronary lesions based on the radiomic signature of perilesional fat. In summary, the findings presented herein indicate that PVR radiomic phenotyping with the aid of the radiomic signature of the present invention can be used to identify both high-risk patients (when measured in a normalized manner around coronary vessels) and high-risk lesions (when applied to specific coronary artery segments or around lesions), which is of great significance to modern CCTA-based risk prediction.

[0414] Surprisingly, the radiomic signature need not be constructed from the radiomic features that are most independently strongly associated with future adverse events. Rather, it is actually advantageous to include radiomic features selected from different co-linear "clusters" of radiomic features, rather than only those radiomic features that are most independently strongly associated with adverse events.

[0415] 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.

[0416] Selected Aspects of the Invention

[0417] The following numbered clauses disclose various aspects of the invention.

[0418] Clause 1. A method for characterizing a perivascular region using medical imaging data of a subject, the method comprising calculating a value of a radiomics signature of the perivascular region using the medical imaging data;

[0419] The radiomic signature is calculated based on measured values ​​of at least two radiomic features of the perivascular region, the measured values ​​of the at least two radiomic features being calculated from the medical imaging data.

[0420] Clause 2. The method of clause 1, wherein the at least two radiomic features are selected from the radiomic features of clusters 1 to 9, wherein the at least two radiomic features are each selected from a different cluster, wherein:

[0421] Cluster 1 consists of short-run high grayscale dominance, high grayscale dominance, high grayscale run dominance, autocorrelation, mean sum, joint mean, and high grayscale region dominance;

[0422] Cluster 2 consists of skewness, skewness LLL, kurtosis, 90th percentile, 90th percentile LLL, median LLL, kurtosis LLL, and median;

[0423] Cluster 3 consists of run entropy, dependency entropy LLL, dependency entropy, regional entropy LLL, run entropy LLL and mean LLL;

[0424] Cluster 4 consists of small area low grayscale advantage, low grayscale area advantage, short run low grayscale advantage, low grayscale run advantage, low grayscale advantage, small dependence low grayscale advantage, grayscale variance LLL (GLSZM), grayscale variance (GLDM), variance, grayscale variance (GLDM), difference variance LLL, grayscale variance LLL (GLRLM), variance LLL, grayscale variance LLL (GLDM), sum of squares, contrast LLL, mean absolute deviation, interquartile range, robust mean absolute deviation, long run low grayscale advantage, difference variance, grayscale variance (GLSZM), normalized inverse moment, mean absolute deviation LLL, sum of squares LLL and contrast;

[0425] Cluster 5 consists of regional entropy, normalized grayscale nonuniformity (GLRLM), normalized grayscale nonuniformity LLL (GLRLM), sum entropy, joint energy, entropy, normalized grayscale nonuniformity (GLDM), joint energy, normalized grayscale nonuniformity LLL (GLDM), uniformity LLL, sum entropy LLL and uniformity;

[0426] Cluster 6 consists of regional entropy HHH, normalized size regional heterogeneity HHH, and small area advantage HHH;

[0427] Cluster 7 consists of intensity, roughness HLL, roughness, roughness LHL, roughness LLL, roughness LLH, roughness HHH, roughness HLH, roughness HHL, and roughness LHH;

[0428] Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, gray variance LLL (GLDM), variance LLL, gray variance LLL (GLRLM), gray variance (GLRLM), robust mean absolute deviation LLL, gray variance (GLDM), variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, gray variance LLL (GLSZM), sum of squares, robust mean absolute deviation, sum entropy, interquartile range, cluster prominence LLL, entropy LLL, 10th percentile LLL, 10th percentile; and

[0429] Cluster 9 consists of size area unevenness LLL, dependence unevenness HLL, grayscale unevenness HLL (GLSZM), grayscale unevenness (GLSZM), run length unevenness HHL, run length unevenness LHL, dependence unevenness LHL, dependence unevenness, run length unevenness HLH, busyness, run length unevenness LLH, dependence unevenness LLH, dependence unevenness LLL, size area unevenness, energy HLL, run length unevenness LHH, size area unevenness HLL, grayscale unevenness LLH (GLSZM), grayscale unevenness LHL (GLSZM), grayscale unevenness LLL (GLSZM), run length unevenness HLL, grayscale unevenness HLH (GLSZM), grayscale unevenness HHL (GLSZM), run length unevenness and run length unevenness HHH.

[0430] Clause 3. The method according to clause 2, wherein

[0431] Cluster 1 consists of short-run high gray-level dominance, high gray-level dominance, high gray-level run dominance, autocorrelation, mean sum, and joint mean;

[0432] Cluster 2 consists of skewness, skewness LLL, kurtosis, 90th percentile, 90th percentile LLL, median LLL, and kurtosis LLL;

[0433] Cluster 3 consists of run entropy, dependency entropy LLL, dependency entropy, and regional entropy LLL;

[0434] Cluster 4 consists of small area low grayscale advantage, low grayscale area advantage, short run low grayscale advantage, low grayscale run advantage, low grayscale advantage, small dependence low grayscale advantage, grayscale variance LLL (GLSZM), grayscale variance (GLDM), variance, grayscale variance (GLDM), and difference variance LLL;

[0435] Cluster 5 consists of regional entropy, normalized gray level non-uniformity (GLRLM), and normalized gray level non-uniformity LLL (GLRLM);

[0436] Cluster 6 consists of the regional entropy HHH and the normalized size regional heterogeneity HHH;

[0437] Cluster 7 consists of intensity, roughness HLL, roughness, roughness LHL, roughness LLL, roughness LLH, and roughness HHH;

[0438] Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, grayscale variance LLL (GLDM), variance LLL, grayscale variance LLL (GLRLM), grayscale variance (GLRLM), robust mean absolute deviation LLL, grayscale variance (GLDM), variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, grayscale variance LLL (GLSZM), sum of squares, robust mean absolute deviation, sum entropy, interquartile range, cluster prominence LLL, entropy LLL, 10th percentile LLL and 10th percentile; and

[0439] Cluster 9 consists of size area unevenness LLL, dependence unevenness HLL, grayscale unevenness HLL (GLSZM), grayscale unevenness (GLSZM), run length unevenness HHL, run length unevenness LHL, dependence unevenness LHL, dependence unevenness, run length unevenness HLH, busyness, run length unevenness LLH, dependence unevenness LLH, dependence unevenness LLL, size area unevenness, energy HLL, run length unevenness LHH, size area unevenness HLL, grayscale unevenness LLH (GLSZM) and grayscale unevenness LHL (GLSZM).

[0440] Clause 4. The method according to clause 2, wherein

[0441] Cluster 1 consists of short-run high gray-level dominance, high gray-level dominance, high gray-level run dominance, autocorrelation, mean sum, and joint mean;

[0442] Cluster 2 consists of skewness, skewness LLL, kurtosis, and 90th percentile;

[0443] Cluster 3 consists of run entropy, dependency entropy LLL, and dependency entropy;

[0444] Cluster 4 consists of small area low grayscale advantage, low grayscale area advantage, short run low grayscale advantage, low grayscale run advantage, low grayscale advantage, and small dependency low grayscale advantage;

[0445] Cluster 5 consists of regional entropy and normalized gray-level non-uniformity (GLRLM);

[0446] Cluster 6 consists of regional entropy HHH;

[0447] Cluster 7 consists of intensity, roughness HLL, roughness, roughness LHL, roughness LLL, and roughness LLH;

[0448] Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, grayscale variance LLL (GLDM), variance LLL, grayscale variance LLL (GLRLM), grayscale variance (GLRLM), robust mean absolute deviation LLL, grayscale variance (GLDM), variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, grayscale variance LLL (GLSZM), sum of squares, robust mean absolute deviation, sum entropy, interquartile range, cluster prominence LLL, entropy LLL, 10th percentile LLL and 10th percentile; and

[0449] Cluster 9 consists of size area unevenness LLL, dependence unevenness HLL, grayscale unevenness HLL (GLSZM), grayscale unevenness (GLSZM), run length unevenness HHL, run length unevenness LHL, dependence unevenness LHL, dependence unevenness, run length unevenness HLH, busyness, run length unevenness LLH, dependence unevenness LLH, dependence unevenness LLL and size area unevenness.

[0450] Clause 5. The method according to clause 2, wherein

[0451] Cluster 1 consists of short-run high grayscale dominance, high grayscale dominance, high grayscale run dominance, and autocorrelation;

[0452] Cluster 2 consists of skewness and skewness LLL;

[0453] Cluster 3 consists of run entropy and dependency entropy LLL;

[0454] Cluster 4 consists of small areas with low grayscale dominance and low grayscale area dominance;

[0455] Cluster 5 consists of regional entropy;

[0456] Cluster 6 consists of regional entropy HHH;

[0457] Cluster 7 consists of intensity;

[0458] Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, grayscale variance LLL (GLDM), variance LLL, grayscale variance LLL (GLRLM), grayscale variance (GLRLM), robust mean absolute deviation LLL, grayscale variance (GLDM), variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, grayscale variance LLL (GLSZM), sum of squares, robust mean absolute deviation, sum entropy, interquartile range, cluster prominence LLL, entropy LLL, 10th percentile LLL and 10th percentile; and

[0459] Cluster 9 consists of large and small area non-uniformity LLL, dependence non-uniformity HLL, grayscale non-uniformity HLL (GLSZM), grayscale non-uniformity (GLSZM), run length non-uniformity HHL and run length non-uniformity LHL.

[0460] Clause 6. The method according to clause 2, wherein

[0461] Cluster 1 consists of short runs with high grayscale dominance, high grayscale dominance, and high grayscale runs with dominance;

[0462] Cluster 2 consists of skewness and skewness LLL;

[0463] Cluster 3 consists of run entropy;

[0464] Cluster 4 consists of small areas with low grayscale dominance and low grayscale area dominance;

[0465] Cluster 5 consists of regional entropy;

[0466] Cluster 6 consists of regional entropy HHH;

[0467] Cluster 7 consists of intensity;

[0468] Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, gray variance LLL (GLDM), variance LLL, gray variance LLL (GLRLM), gray variance (GLRLM), robust mean absolute deviation LLL, gray variance (GLDM), variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, gray variance LLL (GLSZM), sum of squares, robust mean absolute deviation, and sum entropy; and

[0469] Cluster 9 consists of large and small regions of heterogeneity LLL.

[0470] Clause 7. The method according to clause 2, wherein

[0471] Cluster 1 consists of short runs with high grayscale dominance, high grayscale dominance, and high grayscale runs with dominance;

[0472] Cluster 2 consists of skewness, skewness LLL, kurtosis, 90th percentile, median LLL, kurtosis LLL, and median;

[0473] Cluster 3 consists of run entropy, run entropy LLL, and median LLL;

[0474] Cluster 4 consists of small area low gray level dominance, low gray level region dominance, and gray level variance (GLSZM);

[0475] Cluster 5 consists of regional entropy, normalized gray level non-uniformity (GLRLM), and uniformity;

[0476] Cluster 6 consists of regional entropy HHH;

[0477] Cluster 7 consists of intensity;

[0478] Cluster 8 consists of cluster tendency LLL, cluster tendency, mean absolute deviation LLL, grayscale variance LLL (GLDM), variance LLL, grayscale variance LLL (GLRLM), robust mean absolute deviation LLL, grayscale variance LLL (GLSZM), interquartile range LLL, sum entropy LLL, grayscale variance (GLRLM), variance, grayscale variance (GLDM), mean absolute deviation, sum entropy, robust mean absolute deviation, interquartile range, entropy LLL, 10th percentile LLL, 10th percentile, entropy, uniformity LLL, normalized grayscale unevenness LLL (GLDM), normalized grayscale unevenness LLL (GLRLM), root mean square, normalized grayscale unevenness (GLDM), root mean square LLL and long run low grayscale advantage; and

[0479] Cluster 9 consists of LLL, busyness, and LLL size-area unevenness.

[0480] Clause 8. The method of clause 1, wherein the at least two radiomic features are selected from the group consisting of median LLL, mean LLL, median, root mean square LLL, mean, kurtosis, root mean square, runs entropy LLL (GLRLM), uniformity, 90th percentile, normalized grayscale non-uniformity (GLRLM), uniformity LLL, skewness, normalized grayscale non-uniformity LLL (GLRLM), 10th percentile LLL, skewness LLL, 10th percentile, entropy, interquartile range LLL, robust mean absolute deviation LLL, runs entropy (GLRLM), interquartile range, sum entropy (GLCM), normalized grayscale non-uniformity LLL (GLRLM) , dependence unevenness LHL(GLDM), kurtosis LLL, run length unevenness HHL(GLRLM), entropy LLL, robust mean absolute deviation, sum entropy LLL(GLCM), 90th percentile LLL, run entropy HHL(GLRLM), energy, energy LLL, intensity (NGTDM), autocorrelation (GLCM), mean absolute deviation LLL, high gray dominance (GLDM), joint mean (GLCM), mean sum (GLCM), short run high gray dominance (GLRLM), energy HHH, high gray run dominance (GLRLM), run entropy HHH(GLRLM), energy HHL and mean absolute deviation.

[0481] Clause 9. The method according to any one of Clauses 2 to 7, wherein the at least two radiomic features are selected from the radiomic features of clusters 1 to 8.

[0482] Clause 10. The method according to any one of Clauses 2 to 7, wherein the at least two radiomic features are selected from the radiomic features of clusters 1 to 7.

[0483] Clause 11. The method according to any one of Clauses 2 to 7, wherein the at least two radiomic features are selected from the radiomic features of clusters 1 to 6.

[0484] Clause 12. The method according to any one of Clauses 2 to 7, wherein the at least two radiomic features are selected from the radiomic features of clusters 1 to 5.

[0485] Clause 13. The method according to any one of Clauses 2 to 7, wherein the at least two radiomic features are selected from the radiomic features of clusters 1 to 4.

[0486] Clause 14. The method according to any one of Clauses 2 to 7, wherein the at least two radiomic features are selected from the radiomic features of clusters 1 to 3.

[0487] Clause 15. The method according to any one of clauses 2 to 7, wherein the at least two radiomic features are selected from the radiomic features of clusters 1 and 2.

[0488] Clause 16. The method of any one of Clauses 1 to 14, wherein the at least two radiomic features comprise at least three radiomic features.

[0489] Clause 17. The method of any one of Clauses 1 to 13, wherein the at least two radiomic features comprise at least four radiomic features.

[0490] Clause 18. The method of any one of Clauses 1 to 12, wherein the at least two radiomic features comprise at least five radiomic features.

[0491] Clause 19. The method of any one of Clauses 1 to 11, wherein the at least two radiomic features comprise at least six radiomic features.

[0492] Clause 20. The method of any one of Clauses 1 to 10, wherein the at least two radiomic features comprise at least seven radiomic features.

[0493] Clause 21. The method of any one of Clauses 1 to 9, wherein the at least two radiomic features comprise at least eight radiomic features.

[0494] Clause 22. The method of any one of Clauses 1 to 8, wherein the at least two radiomic features comprise at least nine radiomic features.

[0495] Clause 23. The method according to clause 1, wherein the at least two radiomic features include at least six radiomic features, wherein the six radiomic features are short run high grayscale dominance, skewness, run entropy, small area low grayscale dominance, regional entropy HHH and regional entropy.

Claims

1. A method for characterizing a perivascular region including perivascular tissue using medical imaging data of a subject, the method comprising calculating a value of a radiomics signature of the perivascular region using the medical imaging data; wherein the radiomic signature is calculated based on measured values ​​of at least two radiomic features of the perivascular region, the measured values ​​of the at least two radiomic features being calculated based on the medical imaging data; wherein the radiomic signature provides a measure of texture of the perivascular region; and wherein the perivascular region comprises perivascular adipose tissue.

2. The method of claim 1, wherein the radiomic signature predicts cardiovascular risk.

3. The method of claim 1, wherein the radiomic signature predicts the likelihood that the subject will experience a major adverse cardiovascular event. 4 . The method of claim 1 , wherein at least one of the at least two radiomic features is calculated based on a wavelet transform of attenuation values.

5. The method of claim 1, wherein the at least two radiomic features are selected from the radiomic features of clusters 1 to 9, wherein the at least two radiomic features are each selected from a different cluster, wherein: Cluster 1 consists of short-run high grayscale dominance, high grayscale dominance, high grayscale run dominance, autocorrelation, mean sum, joint mean, and high grayscale region dominance; Cluster 2 consists of skewness, skewness LLL, kurtosis, 90th percentile, 90th percentile LLL, median LLL, kurtosis LLL, and median; Cluster 3 consists of run entropy, dependency entropy LLL, dependency entropy, regional entropy LLL, run entropy LLL and mean LLL; Cluster 4 consists of small area low grayscale advantage, low grayscale area advantage, short run low grayscale advantage, low grayscale run advantage, low grayscale advantage, small dependence low grayscale advantage, grayscale variance LLL_GLSZM, grayscale variance_GLDM, variance, grayscale variance_GLDM, difference variance LLL, grayscale variance LLL_GLRLM, variance LLL, grayscale variance LLL_GLDM, sum of squares, contrast LLL, mean absolute deviation, interquartile range, robust mean absolute deviation, long run low grayscale advantage, difference variance, grayscale variance_GLSZM, normalized inverse moment, mean absolute deviation LLL, sum of squares LLL and contrast; Cluster 5 is composed of regional entropy, normalized grayscale nonuniformity_GLRLM, normalized grayscale nonuniformity LLL_GLRLM, sum entropy, joint energy, entropy, normalized grayscale nonuniformity_GLDM, joint energy, normalized grayscale nonuniformity LLL_GLDM, uniformity LLL, sum entropy LLL and uniformity; Cluster 6 consists of regional entropy HHH, normalized size regional heterogeneity HHH, and small area advantage HHH; Cluster 7 consists of intensity, roughness HLL, roughness, roughness LHL, roughness LLL, roughness LLH, roughness HHH, roughness HLH, roughness HHL, and roughness LHH; Cluster 8 consists of cluster tendency LLL, cluster tendency, sum of squares LLL, mean absolute deviation LLL, gray variance LLL_GLDM, variance LLL, gray variance LLL_GLRLM, gray variance_GLRLM, robust mean absolute deviation LLL, gray variance_GLDM, variance, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, gray variance LLL_GLSZM, sum of squares, robust mean absolute deviation, sum entropy, interquartile range, cluster prominence LLL, entropy LLL, 10th percentile LLL, 10th percentile; and Cluster 9 consists of size area unevenness LLL, dependence unevenness HLL, grayscale unevenness HLL_GLSZM, grayscale unevenness_GLSZM, run length unevenness HHL, run length unevenness LHL, dependence unevenness LHL, dependence unevenness, run length unevenness HLH, busyness, run length unevenness LLH, dependence unevenness LLH, dependence unevenness LLL, size area unevenness, energy HLL, run length unevenness LHH, size area unevenness HLL, grayscale unevenness LLH_GLSZM, grayscale unevenness LHL_GLSZM, grayscale unevenness LLL_GLSZM, run length unevenness HLL, grayscale unevenness HLH_GLSZM, grayscale unevenness HHL_GLSZM, run length unevenness and run length unevenness HHH.

6. The method of claim 1, wherein the at least two radiomic features are selected from the group consisting of median LLL, mean LLL, median, root mean square LLL, mean, kurtosis, root mean square, run entropy LLL_GLRLM, uniformity, 90th percentile, normalized grayscale nonuniformity_GLRLM, uniformity LLL, skewness, normalized grayscale nonuniformity LLL_GLRLM, 10th percentile LLL, skewness LLL, 10th percentile, entropy, interquartile range LLL, robust mean absolute deviation LLL, run entropy_GLRLM, interquartile range, and sum entropy_GLCM, normalized grayscale nonuniformity LLL_GL RLM, dependence unevenness LHL_GLDM, kurtosis LLL, run length unevenness HHL_GLRLM, entropy LLL, robust mean absolute deviation, sum entropy LLL_GLCM, 90th percentile LLL, run entropy HHL_GLRLM, energy, energy LLL, intensity_NGTDM, autocorrelation_GLCM, mean absolute deviation LLL, high gray dominance_GLDM, joint mean_GLCM, mean sum_GLCM, short run high gray dominance_GLRLM, energy HHH, high gray run dominance_GLRLM, run entropy HHH_GLRLM, energy HHL and mean absolute deviation. The method according to claim 5 , wherein the at least two radiomic features are selected from the radiomic features of clusters 1 to 6.

8. The method of claim 1, wherein the at least two radiomic signatures include at least six radiomic signatures.

9. The method of any one of claims 1 to 8, further comprising predicting the subject's risk of experiencing a major adverse cardiac event based at least on the calculated value of the radiomic signature.

10. The method according to any one of claims 1 to 8, further comprising determining whether the subject suffers from vascular disease or coronary heart disease based at least on the calculated value of the radiomic signature, or wherein the calculated value of the radiomic signature is used to distinguish unstable from stable coronary artery lesions.

11. A method for deriving a radiomic signature for predicting cardiovascular risk, the method comprising using a radiomics dataset to construct a perivascular radiomic signature for predicting cardiovascular risk, the radiomic signature being calculated based on at least two radiomic features of a perivascular region including perivascular tissue; wherein the data set comprises values ​​of a plurality of radiomic features of perivascular regions, the values ​​of the plurality of radiomic features being obtained from medical imaging data for each of a plurality of individuals, the plurality of individuals comprising a first group of individuals who have reached a clinical endpoint indicative of cardiovascular risk during a subsequent time period after the medical imaging data was collected and a second group of individuals who have not reached a clinical endpoint indicative of cardiovascular risk during the subsequent time period; wherein the radiomic signature is configured to provide a measure of texture of the perivascular region; and wherein the perivascular region comprises perivascular adipose tissue.

12. The method according to claim 11, further comprising: Significant radiomic features are identified from the plurality of radiomic features that are each significantly associated with the clinical endpoint, and the at least two radiomic features are each selected as a different significant radiomic feature or collinear with the different significant radiomic features.

13. The method according to claim 12, further comprising: A subset of the significant radiomic features that are not collinear with each other is identified, and the at least two radiomic features are each selected as belonging to a different radiomic feature of the subset or as being collinear with the different radiomic feature.

14. The method of claim 11, wherein each of the at least two radiomic signatures is selected to be collinear with a corresponding partner radiomic signature that is significantly associated with the clinical endpoint, the partner radiomic signatures of the at least two radiomic signatures being each different from one another.

15. The method of claim 14, wherein each of the partner radiomic signatures is selected to be non-collinear with any other of the other partner radiomic signatures.

16. The method of claim 15, and further wherein at least one of the at least two radiomic signatures is its own partner radiomic signature.

17. A method according to any one of claims 11 to 16, wherein the method comprises identifying a plurality of clusters of radiomic features, each cluster comprising a subset of the plurality of radiomic features, wherein each cluster comprises an original radiomic feature, each of the other radiomic features in the cluster is selected to be collinear with the original radiomic feature, and wherein the at least two radiomic features are each selected from a different cluster. 18 . The method of claim 17 , wherein each of the raw radiomic features is selected to be non-collinear with any of the raw radiomic features of any other cluster in the other clusters.

19. The method of any one of claims 11 to 16, wherein the at least two radiomic features are selected to be non-collinear with each other.

20. The method of any one of claims 11 to 16, wherein the radiomic signature is configured to correlate with the clinical endpoint.

21. The method of claim 20, wherein the radiomic signature is configured to be significantly associated with the clinical endpoint.

22. The method of any one of claims 11 to 16, wherein the at least two radiomic features are selected to be stable.

23. The method of any one of claims 11 to 16, wherein the step of constructing the radiomics signature is performed using a machine learning algorithm.

24. The method of any one of claims 11 to 16, further comprising a system configured to calculate a value of the radiomics signature for a patient.

25. The method of any one of claims 11 to 16, further comprising characterizing the perivascular region of the patient by calculating a value of the radiomic signature for the perivascular region of the patient.

26. The method of any one of claims 1 to 8 and 11-16, wherein the radiomic signature is calculated based on additional radiomic features of the perivascular region in addition to the at least two radiomic features.

27. A system configured to perform the method according to any one of claims 1 to 26.

Images