Radiomics signature of the perivascular region

Abstract

A method for characterizing perivascular regions using medical imaging data of a subject, the method including calculating a value of a radiomics signature of the perivascular region using the medical imaging data. Also disclosed is a method for deriving a radiomics signature for predicting cardiovascular risk, the method including obtaining a radiomics dataset and constructing a perivascular radiomics signature using the radiomics dataset. Also disclosed is a system for performing the aforementioned method. [Selected Figure] Figure 1

Description

Technical Field

[0001] The present invention relates to a method for characterizing a perivascular region using a radiomics signature, and a system therefor. The present invention also relates to a method for deriving such a signature, and a system therefor.

Background Art

[0002] Coronary artery disease (CAD) remains a major contributor to morbidity and mortality despite significant advances in both primary and secondary cardiovascular prevention. Non-invasive diagnostic tests for assessing the presence of CAD in patients presenting with typical or atypical symptoms, such as coronary computed tomography angiography (CCTA), are a mainstay of modern cardiovascular diagnosis, particularly among individuals with a low to moderate pretest probability of CAD. Such techniques have traditionally relied 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 significant number of patients have an elevated residual cardiovascular risk. Residual vascular inflammation in particular is a driver of adverse events and contributes to both atherosclerotic plaque formation and destabilization, but can be difficult to diagnose using conventional assays such as circulating inflammatory biomarkers that are not specific to vascular disease.

[0003] Research into the interpretation of clinical imaging has traditionally relied on subjective, operator-dependent, qualitative assessment of imaged anatomical structures. This approach is useful in a busy clinical environment but ignores the large amount of information contained in all clinical scans.

[0004] WO 2016 / 024128 A1 and WO 2018 / 078395 A1 have found a specific CCTA-based metric, namely a susceptibility and dynamic biomarker of coronary artery inflammation, and subsequently a fat attenuation index (FAI) or perivascular FAI (FAI PVAT ), which reflects the standardized weighted average attenuation of the perivascular region around the human coronary artery and has been identified as a strong and independent predictor of adverse cardiac events. In the presence of vascular inflammation, the release of pro-inflammatory molecules from diseased vessel walls inhibits the differentiation and lipid accumulation in preadipocytes within the perivascular tissue (PVT), resulting in smaller, less differentiated, lipid-free adipocytes. This is associated with a shift in the radiodensity of the PVT in computed tomography (CT) imaging from a more negative (- closer to -190) Hounsfield unit (HU) value to a less negative (- closer to -30) HU value (measured as the attenuation value), and this shift can be captured by the FAI. Although both the biological meaning and clinical value of the FAI have been extensively validated, they are not sufficient to explain the full range of phenotypic variations observed in coronary PVT.

[0005] In a recent study, Kolossvary et al. showed that radiomic features can reliably identify plaques with a high-risk plaque feature napkin-ring signature, and this identification usually depends on the qualitative assessment of the anatomical structure of the plaque and the high expertise of the operator (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 the phenotype of the plaque itself and ignores the PVT and the valuable information that can be obtained from the PVT.

[0006] What is needed is a method or tool that provides prognostic values for cardiovascular risk, such as the presence and extent of CAD, the presence of high-risk plaque features, coronary artery calcium, and recently described FAI, in addition to current CCTA-based risk stratification tools. PVAT SUMMARY OF THE INVENTION

[0007] According to a first aspect of the present invention, there is provided a method for characterizing a perivascular region (e.g., a phenotype of the perivascular region, e.g., composition and / or texture) using medical imaging data of a subject. The method may include calculating values 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 from the medical imaging data or using the medical imaging data.

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

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

[0010] The radiomics signature may be a precursor of cardiovascular risk.

[0011] The radiomics signature may be a precursor of the likelihood that a subject will experience a major adverse cardiovascular event.

[0012] The radiomics signature may be a precursor of the likelihood that a subject will experience a major adverse cardiac-specific cardiovascular event.

[0013] The radiomics signature may indicate cardiovascular health.

[0014] ​A radiomics signature may indicate a vascular disease. For example, a radiomics feature may indicate vascular inflammation.

[0015] At least one of at least two radiomics features may be calculated from a wavelet transform of attenuation values.

[0016] Each of at least two radiomics features may be volume-dependent and / or orientation-dependent.

[0017] At least two radiomics features may be selected from the radiomics features of clusters 1 to 9, and at least two radiomics features are each selected from different clusters. Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, Autocorrelation, Sum Average, Joint Average, and High Gray Level Zone Emphasis. 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, Dependence Entropy LLL, Dependence Entropy, Zone Entropy LLL, Run Entropy LLL, and Mean LLL. Cluster 4 consists of Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, Short Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Low Gray Level Emphasis, Small Dependence Low Gray Level Emphasis, Gray Level Variance LLL(GLSZM), Gray Level Variance(GLDM), Variance, Gray Level Variance(GLDM), Difference Variance LLL, Gray Level Variance LLL(GLRLM), Variance LLL, Gray Level Variance LLL(GLDM), Sum of Squares, Contrast LLL, Mean Absolute Deviation, Interquartile Range, Robust Mean Absolute Deviation, Long Run Low Gray Level Emphasis, Difference Variance, Gray Level Variance(GLSZM), Inverse Difference Moment Normalized, Mean Absolute Deviation LLL, Sum of Squares LLL, and Contrast, Cluster 5 consists of Zone Entropy, Gray Level Non Uniformity Normalized (GLRLM), Gray Level Non Uniformity Normalized LLL (GLRLM), Sum Entropy, Joint Energy, Entropy, Gray Level Non Uniformity Normalized (GLDM), Joint Energy, Gray Level Non Uniformity Normalized LLL (GLDM), Uniformity LLL, Sum Entropy LLL, and Uniformity (Zone Entropy, Gray Level Non Uniformity Normalized(GLRLM), Gray Level Non Uniformity Normalized LLL(GLRLM), Sum Entropy, Joint Energy, Entropy, Gray Level Non Uniformity Normalized(GLDM), Joint Energy, Gray Level Non Uniformity Normalized LLL(GLDM), Uniformity LLL, Sum Entropy LLL, and Uniformity), Cluster 6 consists of Zone Entropy HHH, Size Zone Non Uniformity Normalized HHH, and Small Area Emphasis HHH (Zone Entropy HHH, Size Zone Non Uniformity Normalized HHH, and Small Area Emphasis HHH), Cluster 7 consists of Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, Coarseness LLH, Coarseness HHH, Coarseness HLH, Coarseness HHL, and Coarseness LHH (Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, Coarseness LLH, Coarseness HHH, Coarseness HLH, Coarseness HHL, and Coarseness LHH), Cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level Variance LLL (GLSZM), Sum of Squares, Robust Mean Absolute Deviation, Sum Entropy, Interquartile Range, Cluster Prominence LLL, Entropy LLL, 10th Percentile LLL, 10th Percentile (Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance(GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance(GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level Variance LLL (GLSZM), Sum of Squares, Robust Mean Absolute Deviation, Sum Entropy, Interquartile Range, Cluster Prominence LLL, Entropy LLL, 10th Percentile LLL, 10th Percentile), Cluster 9 has size zone non-uniformity LLL, dependence non-uniformity HLL, gray level non-uniformity HLL (GLSZM), gray level non-uniformity (GLSZM), run length non-uniformity HHL, run length non-uniformity LHL, dependence non-uniformity LHL, dependence non-uniformity, run length non-uniformity HLH, busyness, run length non-uniformity LLH, dependence non-uniformity LLH, dependence non-uniformity LLL, size zone non-uniformity, energy HLL, run length non-uniformity LHH, size zone non-uniformity HLL, gray level non-uniformity LLH (GLSZM), gray level non-uniformity LHL (GLSZM), gray level non-uniformity LLL (GLSZM), run length non-uniformity HLL, gray level nonIt consists of Uniformity HLH(GLSZM), Gray Level Non Uniformity HHL(GLSZM), Run Length Non Uniformity, and Run Length Non Uniformity HHH.

[0018] Clusters 1 to 9 may instead be defined as follows: Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, Autocorrelation, Sum Average, and Joint Average. Cluster 2 consists of Skewness, Skewness LLL, Kurtosis, 90th Percentile, 90th Percentile LLL, Median LLL, and Kurtosis LLL. Cluster 3 consists of Run Entropy, Dependence Entropy LLL, Dependence Entropy, and Zone Entropy LLL. Cluster 4 consists of Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, Short Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Low Gray Level Emphasis, Small Dependence Low Gray Level Emphasis, Gray Level Variance LLL (GLSZM), Gray Level Variance (GLDM), Variance, Gray Level Variance (GLDM), and Difference Variance LLL, Cluster 5 consists of Zone Entropy, Gray Level Non Uniformity Normalized (GLRLM), and Gray Level Non Uniformity Normalized LLL (GLRLM) (Zone Entropy, Gray Level Non Uniformity Normalized (GLRLM), and Gray Level Non Uniformity Normalized LLL (GLRLM)), Cluster 6 consists of Zone Entropy HHH and Size Zone Non Uniformity Normalized HHH (Zone Entropy HHH and Size Zone Non Uniformity Normalized HHH), Cluster 7 consists of Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, Coarseness LLH, and Coarseness HHH (Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, Coarseness LLH, and Coarseness HHH), Cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level 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 (Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance(GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance(GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level 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), Cluster 9 consists of Size Zone Non Uniformity LLL, Dependence Non Uniformity HLL, Gray Level Non Uniformity HLL(GLSZM), Gray Level Non Uniformity(GLSZM), Run Length Non Uniformity HHL, Run Length Non Uniformity LHL, Dependence Non Uniformity LHL, Dependence Non Uniformity, Run Length Non Uniformity HLH, Busyness, Run Length Non Uniformity LLH, Dependence Non Uniformity LLH, Dependence Non Uniformity LLL, Size Zone Non Uniformity, Energy HLL, Run Length Non Uniformity LHH, Size Zone Non Uniformity HLL, Gray Level Non Uniformity LLH(GLSZM), and Gray Level Non Uniformity LHL(GLSZM).

[0019] Clusters 1 - 9 may instead be defined as follows: Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, Autocorrelation, Sum Average, and Joint Average, Cluster 2 consists of Skewness, Skewness LLL, Kurtosis, and 90th Percentile, Cluster 3 consists of Run Entropy, Dependence Entropy LLL, and Dependence Entropy, Cluster 4 consists of Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, Short Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Low Gray Level Emphasis, and Small Dependence Low Gray Level Emphasis, Cluster 5 consists of Zone Entropy, and Gray Level Non Uniformity Normalized (GLRLM), Cluster 6 consists of Zone Entropy HHH, Cluster 7 consists of Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, and Coarseness LLH, Cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL(GLDM), Variance LLL, Gray Level Variance LLL(GLRLM), Gray Level Variance(GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance(GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level 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, Cluster 9 consists of Size Zone Non Uniformity LLL, Dependence Non Uniformity HLL, Gray Level Non Uniformity HLL (GLSZM), Gray Level Non Uniformity (GLSZM), Run Length Non Uniformity HHL, Run Length Non Uniformity LHL, Dependence Non Uniformity LHL, Dependence Non Uniformity, Run Length Non Uniformity HLH, Busyness, Run Length Non Uniformity LLH, Dependence Non Uniformity LLH, Dependence Non Uniformity LLL, and Size Zone Non Uniformity.

[0020] Clusters 1 - 9 may instead be defined as follows: Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, and Autocorrelation. Cluster 2 consists of Skewness, and Skewness LLL. Cluster 3 consists of Run Entropy, and Dependence Entropy LLL. Cluster 4 consists of Small Area Low Gray Level Emphasis, and Low Gray Level Zone Emphasis, Cluster 5 consists of Zone Entropy, Cluster 6 consists of Zone Entropy HHH, Cluster 7 consists of Strength, Cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL(GLDM), Variance LLL, Gray Level Variance LLL(GLRLM), Gray Level Variance(GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance(GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level 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 (Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL(GLDM), Variance LLL, Gray Level Variance LLL(GLRLM), Gray Level Variance(GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance(GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level 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). Cluster 9 consists of Size Zone Non Uniformity LLL, Dependence Non Uniformity HLL, Gray Level Non Uniformity HLL (GLSZM), Gray Level Non Uniformity (GLSZM), Run Length Non Uniformity HHL, and Run Length Non Uniformity LHL.

[0021] Clusters 1 to 9 may alternatively be defined as follows: Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, and High Gray Level Run Emphasis. Cluster 2 consists of Skewness and Skewness LLL. Cluster 3 consists of Run Entropy. Cluster 4 consists of Small Area Low Gray Level Emphasis and Low Gray Level Zone Emphasis. Cluster 5 consists of Zone Entropy. Cluster 6 consists of Zone Entropy HHH. Cluster 7 consists of Strength. Cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level Variance LLL (GLSZM), Sum of Squares, Robust Mean Absolute Deviation, and Sum Entropy. Cluster 9 consists of Size Zone Uniformity LLL.

[0022] Clusters 1 to 9 may instead be defined as follows: Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, and High Gray Level Run Emphasis. Cluster 2 consists of skewness, skewness LLL, kurtosis, 90th percentile, median LLL, kurtosis LLL, and median (Skewness, Skewness LLL, Kurtosis, 90th Percentile, Median LLL, Kurtosis LLL, and Median). Cluster 3 consists of run entropy, run entropy LLL, and mean LLL (Run Entropy, Run Entropy LLL, and Mean LLL). Cluster 4 consists of small area low gray level emphasis, low gray level zone emphasis, and gray level variance (GLSZM) (Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, and Gray Level Variance (GLSZM)). Cluster 5 consists of zone entropy, gray level non-uniformity normalized (GLRLM), and uniformity (Zone Entropy, Gray Level Non Uniformity Normalized (GLRLM), and Uniformity). Cluster 6 consists of zone entropy HHH (Zone Entropy HHH). Cluster 7 consists of strength (Strength). Cluster 8 is characterized by Cluster Tendency LLL, Cluster Tendency, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance LLL (GLSZM), Interquartile Range LLL, Sum Entropy LLL, Gray Level Variance (GLRLM), Variance, Gray Level Variance (GLDM), Mean Absolute Deviation, Sum Entropy, Robust Mean Absolute Deviation, Interquartile Range, Entropy LLL, 10th Percentile LLL, 10th Percentile, Entropy, Uniformity LLL, Gray Level Non Uniformity Normalized LLL (GLDM), Gray Level Non Uniformity Normalized LLL (GLRLM), Root Mean Squared, Gray Level Non Uniformity Normalized (GLDM), Root Mean Squared LLL, and Longconsisting of Run Low Gray Level Emphasis), Cluster 9 consists of Size Zone Non Uniformity LLL, Busyness, and Size Zone Non Uniformity.

[0023] At least two radiomics features are the median LLL, mean LLL, median, root mean squared LLL, mean, kurtosis, root mean squared, run entropy LLL (GLRLM), uniformity, 90th percentile, gray level non-uniformity normalized (GLRLM), uniformity LLL, skewness, gray level non-uniformity normalized LLL (GLRLM), 10th percentile LLL, skewness LLL, 10th percentile, entropy, interquartile range LLL, robust mean absolute deviation LLL, run entropy (GLRLM), interquartile range, sum entropy (GLCM), gray levelNon-Uniformity Normalized LLL (GLRLM), Dependence Non-Uniformity LHL (GLDM), Kurtosis LLL, Run Length Non-Uniformity HHL (GLRLM), Entropy LLL, Robust Mean Absolute Deviation, Sum Entropy LLL (GLCM), 90th Percentile LLL, Run Entropy HHL (GLRLM), Energy, Energy LLL, Strength (NGTDM), Autocorrelation (GLCM), Mean Absolute Deviation LLL, High Gray Level Emphasis (GLDM), Joint Average (GLCM), Sum Average (GLCM), Short Run High Gray Level Emphasis (GLRLM), Energy HHH, High Gray Level Run Emphasis (GLRLM), Run Entropy HHH (GLRLM), Energy HHL, and Mean Absolute Deviation) can be selected.

[0024] At least two radiomics features can be selected from the radiomics features of clusters 1 to 8. At least two radiomics features can be selected from the radiomics features of clusters 1 to 7. At least two radiomics features can be selected from the radiomics features of clusters 1 to 6. At least two radiomics features can be selected from the radiomics features of clusters 1 to 5. At least two radiomics features can be selected from the radiomics features of clusters 1 to 4. At least two radiomics features can be selected from the radiomics features of clusters 1 to 3. At least two radiomics features can be selected from the radiomics features of clusters 1 and 2.

[0025] At least two radiomics features may include at least three radiomics features. At least two radiomics features may include at least four radiomics features. At least two radiomics features may include at least five radiomics features. At least two radiomics features may include at least six radiomics features. At least two radiomics features may include at least seven radiomics features. At least two radiomics features may include at least eight radiomics features. At least two radiomics features may include at least nine radiomics features.

[0026] At least two radiomics features may include six radiomics features, and the six radiomics features are Short Run High Gray Level Emphasis, Skewness, Run Entropy, Small Area Low Gray Level Emphasis, Zone Entropy HHH, and Zone Entropy.

[0027] At least two radiomics features may include nine radiomics features, and the nine radiomics features are Short Run High Gray Level Emphasis, Skewness, Run Entropy, Small Area Low Gray Level Emphasis, Zone Entropy HHH, Zone Entropy, Strength, Cluster Tendency LLL, and Size Zone Non Uniformity.

[0028] A radiomics signature may include a weighted sum of the values of at least two radiomics features. The radiomics signature may be linearly related to the weighted sum of the values of at least two radiomics features.

[0029] Medical imaging data may include the attenuation value of each of a plurality of voxels corresponding to at least a perivascular region. The plurality of voxels may also correspond to a blood vessel with a perivascular region disposed around it.

[0030] This method may further include identifying a perivascular region (or substance, tissue, or type of perivascular tissue, such as adipose tissue) using medical imaging data. The perivascular region may be identified as all voxels of the medical imaging data having attenuation (or radiation density) values within a given range of attenuation values and / or within a given radial distance from an outer vessel wall (i.e., the blood vessel adjacent to the perivascular tissue). The given range of attenuation values may be, for example, about -190 to about -30 Hounsfield units when the perivascular region corresponds to perivascular adipose tissue. The given range of attenuation values may be, for example, about -30 to about +30 Hounsfield units when the perivascular region corresponds to water. The given radial distance may be a distance related to one or more dimensions of the adjacent blood vessel. The given radial distance may be equal to the diameter of the adjacent blood vessel. The given radial distance may be a fixed value, such as 5 mm.

[0031] The perivascular region may be adjacent to or surround a coronary artery, carotid artery, aorta, or any other artery in the human body. For example, the perivascular region may be adjacent to or surround a proximal and / or intermediate right coronary artery.

[0032] This method may further include segmenting the perivascular region. This method may further include calculating the values of radiomics features from the segmented perivascular region.

[0033] The value of each of at least two radiomics features can be calculated from the original attenuation value, the binned attenuation value, or the wavelet transform of the attenuation value.

[0034] This method may further include predicting the risk that a subject will experience a major adverse cardiac event based on the calculated value of the radiomics signature. This method may further include predicting the risk that a subject will experience a major adverse cardiac event specific to the heart based on the calculated value of the radiomics signature.

[0035] This method may further include determining whether a subject has a vascular disease based on the calculated value of the radiomics signature. The vascular disease may be selected from the group consisting of atherosclerosis, vascular calcification, intimal hyperplasia, vascular aneurysm, and vascular inflammation.

[0036] This method may further include determining whether a subject has coronary heart disease based on the calculated value of the radiomics signature.

[0037] The calculated value of the radiomics signature can be used to distinguish between unstable coronary lesions and stable coronary lesions. The perivascular region can be the region near the lesion.

[0038] According to a second aspect of the present invention, a method for deriving a radiomics signature for predicting cardiovascular risk is provided. This method may include constructing a perivascular radiomics signature or score for predicting cardiovascular risk using a dataset, such as a radiomics dataset. The radiomics signature can be calculated based on at least two radiomics features of the perivascular region. The dataset may include measured values of a plurality of radiomics features of the perivascular region obtained from medical imaging data of each of a plurality of individuals. The plurality of individuals may include a first group of individuals who have reached a clinical endpoint indicating cardiovascular risk and a second group of individuals who have not reached a clinical endpoint indicating cardiovascular risk.

[0039] Individuals in the first group may reach a clinical endpoint indicating a cardiovascular risk within a subsequent period after the medical imaging data has been collected, and individuals in the second group may not reach a clinical endpoint indicating a cardiovascular risk within a subsequent period after the medical imaging data has been collected.

[0040] Each of at least two radiomics features may be selected to be collinear (or correlated) with a corresponding partner radiomics feature significantly (i.e., statistically significantly) associated with the clinical endpoint, for example, as determined or calculated from a dataset. The partner radiomics features of at least two radiomics features may each be different from one another.

[0041] Each of at least two radiomics features may be selected to be either a different significant radiomics feature significantly associated with the clinical endpoint (as determined from the dataset, i.e., identified from the dataset as being significantly associated with the clinical endpoint) or collinear with these different significant radiomics features. Thus, the method may further include identifying significant radiomics features from among a plurality of radiomics features each significantly associated with the clinical endpoint, and at least two radiomics features are each selected to be either different significant radiomics features or collinear with different significant radiomics features.

[0042] The significant radiomics features may be selected to not be collinear with one another. For example, the method may further include identifying a subset of a plurality of radiomics features that are not collinear with one another (e.g., a subset of significant radiomics features), and at least two radiomics features are each selected to be either different radiomics features belonging to the subset or collinear with different radiomics features belonging to the subset.

[0043] Each of the partner radiomics features can be selected so as not to be collinear with any of the other partner radiomics features.

[0044] At least one of the at least two radiomics features can be its own partner radiomics feature.

[0045] Each of the at least two radiomics features can be selected to be significantly associated with a clinical endpoint, for example, as determined or calculated from a dataset.

[0046] This method can include identifying a plurality of clusters of radiomics features. Each cluster can include a subset of the plurality of radiomics features. Each cluster can include the original radiomics features selected such that each of the other radiomics features within that cluster is collinear, for example, as determined or calculated from a dataset. At least two radiomics features can each be selected from different clusters.

[0047] Each of the original radiomics features can be selected so as not to be collinear with any of the original radiomics features of any of the other clusters, for example, as determined or calculated from a dataset.

[0048] Each of the radiomics features within each cluster can be selected to be collinear with all of the other radiomics features within the same cluster, for example, as determined or calculated from a dataset.

[0049] Each of the original radiomics features can be selected to be significantly associated with a clinical endpoint, for example, as determined or calculated from a dataset.

[0050] Each of the original radiomics features can be selected to be most strongly associated with the clinical endpoint of all radiomics features within its cluster, for example, as determined or calculated from a dataset.

[0051] At least two radiomics features can be selected to not be collinear with each other, for example, as determined or calculated from a dataset.

[0052] Two radiomics features can be identified as collinear if those radiomics features correlate to an extent greater than a correlation threshold.

[0053] Collinearity between radiomics features can be calculated using Spearman's rho coefficient. Alternatively, collinearity between radiomics features may be calculated using other measures of pairwise correlation, such as Pearson's correlation coefficient (Pearson's r).

[0054] The correlation threshold can be at least about |0.75|, for example, about |rho| = 0.75.

[0055] A radiomics feature can be identified as significantly associated with a clinical endpoint if the radiomics feature is associated with a clinical endpoint that exceeds a significance threshold, for example, as determined or calculated from a dataset. The significance threshold can be at least about α = 0.05, for example, about α = 0.05.

[0056] The association of radiomics features with a clinical endpoint can be calculated based on receiver operating characteristic (ROC) curve analysis, using in particular the area under the curve (AUC) measurement (i.e., C statistic), as will be readily understood by those skilled in the art.

[0057] A Bonferroni correction can be applied to the significance threshold.

[0058] For example, principal component analysis of a plurality of radiomics features can be performed using a dataset (specifically, with respect to the values of radiomics features of a plurality of individuals). The Bonferroni correction can be based on the number of principal components that account for a given amount of the observed variation, as determined from the principal component analysis. The given amount of the observed variation can be at least about 99.5%, for example, about 99.5%.

[0059] A radiomics signature can be constructed to correlate with a clinical endpoint. A radiomics signature can be constructed to be significantly associated with a clinical endpoint.

[0060] A radiomics signature can be identified as being significantly associated with a clinical endpoint if, for example, as determined or calculated from a dataset, the radiomics signature is associated with a clinical endpoint that exceeds a significance threshold. The significance threshold can be at least about 0.05, for example, about α = 0.05.

[0061] The dataset can be split into data of an individual's training cohort and data of an individual's validation cohort. The step of constructing a radiomics signature can include deriving the signature using at least the data of the training cohort. The step of constructing a radiomics signature can include validating the signature using the data of the validation cohort.

[0062] A radiomics feature can be identified as being significantly associated with a clinical endpoint only if the radiomics feature is significantly associated with the clinical endpoint in both cohorts. A radiomics signature can be identified as being significantly associated with a clinical endpoint if the radiomics signature is significantly associated with the clinical endpoint in the training cohort.

[0063] Each of at least two radiomics features may be volume-independent. Each of at least two radiomics features may be orientation-independent. Each of a plurality of radiomics features may be volume-independent and / or orientation-independent.

[0064] Any volume-independent and / or orientation-independent radiomics features may be removed from the plurality of radiomics features before selecting at least two radiomics features.

[0065] At least two radiomics features may be selected to be stable, such as being determined or calculated from a dataset. For example, at least two radiomics features may be selected from those identified as being stable, such as being determined or calculated from a dataset.

[0066] All unstable features may be removed from the plurality of radiomics features. For example, all unstable features may be removed from the plurality of radiomics features before at least two radiomics features are selected.

[0067] A radiomics feature may be identified as unstable if the intraclass correlation coefficient (ICC) (calculated for repeated measurements or scans) of that radiomics feature is less than a stability threshold. The stability threshold may be at least about 0.9, such as 0.9. The intraclass correlation coefficient may be calculated based on the Z-score (or standard score, i.e., represented by a number of standard deviations from the mean) of the radiomics feature.

[0068] The step of constructing a radiomics signature may include refining the contribution of at least two radiomics features to the radiomics signature to increase the association or correlation of the radiomics signature with a clinical endpoint. The radiomics signature may be constructed to be significantly associated with a clinical endpoint.

[0069] The association of a radiomics signature with a clinical endpoint can be calculated based on receiver operating characteristic (ROC) curve analysis, particularly using the area under the curve (AUC) measurement (i.e., C-statistic), as readily understood by those skilled in the art.

[0070] The step of constructing a radiomics signature can be performed using a machine learning algorithm.

[0071] The step of constructing a radiomics signature can be performed using leave-p-out cross-validation.

[0072] The step of constructing a radiomics signature can be performed using leave-one-out cross-validation.

[0073] The step of constructing a radiomics signature can be performed using elastic net regression.

[0074] A radiomics signature can include a weighted sum of at least two radiomics features.

[0075] A radiomics signature can be linearly related to a weighted sum of at least two radiomics features.

[0076] The step of constructing a radiomics signature can include adjusting the relative weights of each of at least two radiomics features to increase the association or correlation of the radiomics signature with a clinical endpoint.

[0077] A radiomics signature can be constructed to provide a measure of the texture of the perivascular region.

[0078] At least one of the at least two radiomics features can provide a measure of the texture of the perivascular region. For example, each of the at least two radiomics features can provide a measure of the texture of the perivascular region (i.e., each of the at least two radiomics features can be a texture statistic).

[0079] The clinical endpoint can be a composite endpoint of serious adverse cardiac events. The clinical endpoint can be a composite endpoint of cardiac-specific serious adverse cardiac events.

[0080] The method of the present invention may also include a step of calculating radiomics features from medical imaging data.

[0081] The radiomics signature of the present invention can also be calculated based on additional radiomics features in addition to the at least two radiomics features mentioned above. Thus, it can be said that the radiomics signature is calculated based on a plurality of radiomics features, and the plurality of radiomics features can include at least two radiomics features. For example, a method for deriving a radiomics signature for predicting cardiovascular risk may include constructing a perivascular radiomics signature or score for predicting cardiovascular risk using a dataset, particularly a radiomics dataset. The radiomics signature can be calculated based on a (second) plurality of perivascular radiomics features (i.e., radiomics features of the perivascular region). The dataset can include measured values of a (first) plurality of perivascular radiomics features of the perivascular region obtained from the medical imaging data of each of a plurality of individuals. The plurality of individuals can include a first group of individuals who have reached a clinical endpoint indicating cardiovascular risk and a second group of individuals who have not reached a clinical endpoint indicating cardiovascular risk. The second plurality of perivascular radiomics features can be selected from among the first plurality of perivascular radiomics features, particularly to provide a radiomics signature for predicting cardiovascular risk as determined from the dataset or as determined using the dataset, for example, using a machine learning algorithm. Thus, the radiomics signature can be calculated based on additional radiomics features (e.g., selected from the (first) plurality of radiomics features) in addition to at least two radiomics features.

[0082] This method may further include configuring a system for calculating a value of a radiomics signature for a patient. For example, this method may further include configuring a system for characterizing the perivascular region of a patient or subject by calculating a value of a derived radiomics signature for the patient or subject. The system may be configured to calculate a value of the derived radiomics 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 a value of the derived radiomics signature using or at least based on values of at least two (or a second plurality) of the radiomics features of the perivascular region of the patient or subject.

[0083] Thus, this method may be for configuring a system for deriving a perivascular radiomics signature and using the derived radiomics signature to characterize the perivascular region of a patient.

[0084] The system may be configured to receive, as input, medical imaging data or values of at least two (or a second plurality) of radiomics features. The system may be configured to output (e.g., display) a calculated value of the radiomics signature or a value based on the calculated value of the radiomics signature. The system may be configured to output a display of the cardiovascular risk and / or vascular health of the patient. The system may be configured to output a display of the risk that the patient will experience an adverse cardiovascular event. The system may be a computer system.

[0085] This method may include providing instructions for configuring a system for calculating a value of a derived radiomics signature for a patient or subject.

[0086] This method may further include calculating the value of the radiomics signature derived for the perivascular region of a patient or subject. For example, this method may further include characterizing the perivascular region of a patient or subject by calculating the value of the derived radiomics signature. The value of the derived radiomics signature may be calculated based on, or using, the medical imaging data of at least the perivascular region of a patient or subject. The value of the derived radiomics signature may be calculated using, or at least based on, the values of at least two (or a second plurality of) radiomics features of the perivascular region of a patient or subject.

[0087] Thus, this method may be for deriving a perivascular radiomics signature and using the derived radiomics signature to characterize the perivascular region.

[0088] According to a third aspect of the present invention, there is provided a system configured to perform any of the methods as described above. The system may be a computer system. The system may include a processor configured to perform the steps of this method. The system may include a memory loaded with executable instructions for performing the steps of this method.

[0089] According to a fourth aspect of the present invention, there is provided the use of a perivascular radiomics signature for any of the above purposes, for example, for characterizing a perivascular region, for detecting a vascular disease, or for predicting a cardiovascular risk. The perivascular radiomics signature may be calculated based on the measured values of a plurality of perivascular radiomics features of the perivascular region.

[0090] The medical imaging data may be radiation data. The medical imaging data may be computed tomography data.

[0091] The perivascular region can be or can include perivascular tissue, such as perivascular adipose tissue. The perivascular region can also include water and / or other soft tissue structures within the perivascular region.

Brief Description of the Drawings

[0092] The present invention will be described with reference to the following attached drawings.

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Mode for Carrying Out the Invention

[0093] The inventors have found that a PVR radiomics signature (also known as a "score" or "index") calculated based on two or more radiomic features of the perivascular region (PVR) adds an incremental value beyond traditional risk factors, predicts future adverse cardiovascular events, and establishes a CCTA risk classification tool for assessing cardiovascular health and risk, and further aids in the detection of general vascular inflammation, local plaque inflammation, and the presence of unstable coronary lesions. Thus, the PVR radiomics signature of the present invention is preferably calculated based on two or more radiomic features of the PVR, evaluates cardiovascular or vascular health, predicts the risk of future adverse cardiovascular events in a patient, identifies or diagnoses coronary disease or coronary heart disease, and provides a tool for characterizing the PVR, such as perivascular tissue such as perivascular adipose tissue (PVAT), for the purpose of identifying unstable coronary lesions or vascular inflammation, such as caused by local plaques.

[0094] The PVR radiomics signature of the present invention can be used alone to characterize a PVR or to provide diagnostic or prognostic information, or this PVR radiomics signature can be combined with existing models such as FAI PVAT , existing models such as the Duke prognostic CAD index, and / or other conventional models including demographic and risk factors such as the presence of coronary artery calcium, high-risk plaque features, and / or epicardial adipose tissue (EAT) volume.

[0095] The present invention utilizes the fact that the coronary artery wall and adjacent PVR, particularly the tissues within the PVR such as adipose tissue, interact bidirectionally. Thus, angiogenic phenotypic changes in coronary PVR can function as a sensor for underlying diseases even in the absence of visible coronary lesions. In particular, the present invention utilizes the effect of this interaction on the texture of the PVR (e.g., spatial heterogeneity or variability), and thus, the radiomics signature of the present invention can be constructed to provide a measure of the texture of the PVR. Thus, the radiomics signature of the present invention can also be referred to as a perivascular texture index (PTI). However, the radiomics-based approach used to construct the signature of the present invention is not specific to constructing a radiomics signature that measures texture, and this approach is primarily important, the prognostic value of the resulting signature. Thus, it is not essential for a radiomics signature to measure texture in order to be an effective prognostic or diagnostic tool.

[0096] PVR refers to the region or volume adjacent to a blood vessel. The PVR can be the region or volume of the perivascular tissue (PVT), or can include the PVT, or can consist of the PVT. The perivascular tissue is the tissue located adjacent to the blood vessel. The tissue is a complex biological structure and can include cells (such as adipocytes, neurons, etc.), as well as extracellular structures and materials (such as water) that can occupy the intercellular space. In particular, the PVT can include or consist of perivascular adipose tissue (PVAT), and thus, the PVR can alternatively be referred to as the region or volume of the PVAT.

[0097] The present invention utilizes a radiomics approach. Radiomics is a field of imaging in which a large amount of quantitative information is extracted from imaging data using data feature evaluation algorithms. The resulting features, referred to as radiomics features, range from simple volume measurement statistics, shape-related statistics, or primary statistics (such as mean or median attenuation) to texture of segmented volumes or regions, and second-order and higher-order statistics that describe the spatial relationships of voxels having similar or different attenuation values. Such features can identify imaging patterns of significant clinical value that are not recognizable by the naked eye and have the potential to maximize the diagnostic yield of non-invasive PVR phenotypes.

[0098] The signature of the present invention is derived and calculated based on radiomics features, for example, features extracted from medical imaging data. In particular, the medical imaging data from which the radiomics features are extracted corresponds to the perivascular region (PVR), such as the coronary perivascular adipose tissue (PVAT) of the coronary artery, and optionally also corresponds to the blood vessel itself and / or other tissues adjacent to or surrounding the PVR. The medical imaging data typically includes radiation density (or attenuation) values, usually in Hounsfield units (HU), for a plurality of voxels of the relevant region, in this case the PVR, and optionally also for adjacent tissues.

[0099] Medical imaging data is preferably computed tomography (CT) data, such as coronary computed tomography angiography (CCTA), but alternatively other forms of medical imaging data that provide attenuation (or radiation density) data for voxels of the imaged region, such as three-dimensional computed laminography data (e.g., radiomics imaging data), may be used. Typically, the medical imaging data used in the present invention is three-dimensional imaging data. Throughout the following, when referring to CCTA or another medical imaging technique, it should be understood that other suitable medical imaging techniques may alternatively be used.

[0100] The PVR may include only voxels having a radiation density (or attenuation) that is within a given or predetermined range and / or within a delineated region, e.g., within a given or predetermined radial distance from an 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 or fixed value, such as about 5 mm. Alternatively, the PVR may be identified by manual contouring or delineation in combination with applying a radiation density or attenuation mask such that only voxels having a radiation density within an arbitrarily and specified range and within the delineated region are included. For example, an operator may identify the PVR through examination of data, such as a CT image. The PVR may include only voxels having a radiation density within a Hounsfield unit range of from about -190 HU to about +30 HU. For example, the PVR may include only voxels having a radiation density within a Hounsfield unit range of from about -190 HU to about -30 HU. This attenuation value range generally corresponds to the radiation density of perivascular adipose tissue (PVAT). However, other ranges, such as a Hounsfield unit range of from about -30 to about +30, which generally corresponds to the radiation density of water, may be used. In particular, the PVR may be identified as all voxels having a radiation density within a Hounsfield unit range of from about -190 HU to about -30 HU and located within a radial distance from an adjacent outer vessel wall that is approximately equal to the diameter of the adjacent vessel.

[0101] Radiomics features, and thus also radiomics signatures, can be calculated for specific blood vessels, for example, coronary blood vessels such as the coronary arteries. The proximal and middle right coronary artery (RCA) (segments 1 and 2 according to the anatomical classification of the American Heart Association, defined in Austen WG, Edwards JE, Frye RL, et al. A reporting system on patients evaluated for coronary artery disease. Report of the Ad Hoc Committee for Grading of Coronary Artery Disease, Council on Cardiovascular Surgery, American Heart Association. Circulation 1975;51(4 Suppl):5 - 40) are particularly suitable due to their straight course and the absence of large branches. Thus, the PVR can be located adjacent to a specific blood vessel.

[0102] The PVR can be segmented before calculating the radiomics features, and the radiomics features can be calculated from the segmented data. The segmented volume or region corresponds to the PVR, and segmentation can remove data corresponding to voxels outside the PVR. Thus, segmentation can be achieved, as described above, by identifying the PVR and then removing any voxels, for example, voxels corresponding to surrounding or adjacent tissue voxels, from the data identified as not being part of the PVR. Segmentation can be performed by placing a three - dimensional sphere having a diameter equal to twice the given distance from the outer vessel wall, on which the PVR can be identified, along the centerline of the blood vessel on consecutive slices within the outer vessel wall, to the diameter of the blood vessel. For example, if the PVR is identified as being located within a radial distance equal to the diameter of an adjacent blood vessel from the outer vessel wall, the sphere will have a diameter equal to three times the diameter of the adjacent blood vessel. The segmented PVR can be extracted and used to calculate radiomics features.

[0103] The calculation of radiomics features from medical imaging data can be performed using a computer program or software. For this purpose, there are various commercially available software packages, such as 3D Slicer. Radiomics features can be shape-related statistics, first-order statistics, or texture statistics (e.g., second- and higher-order statistics). Shape-related and first-order radiomics features can be calculated using the original radiation density (HU) values of 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 neighborhood gray-tone difference matrix [NGTDM], see FIGS. 1A and Tables R1-R7), the PVR voxel radiation density or attenuation value is preferably discretized into a plurality of bins, preferably 16 bins of preferably equal width (e.g., a width of 10 HU), to reduce noise while enabling sufficient resolution to detect biologically significant spatial variations in PVR attenuation. Discretization into 16 bins is recommended as an optimal approach to increase the signal-to-noise ratio of the image for radiomics analysis. However, more or fewer than 16 bins of discretization are also possible. To enforce symmetric, rotation-invariant results, some or all of the radiomics features, particularly texture statistics (such as GLCM), can be calculated in all (orthogonal) directions and then averaged (e.g., using the mean value or other average of the individually calculated values of the features in each of the four directions).

[0104] Some or all of the radiomics features, particularly those related to first-order and texture-based statistics, can also be calculated for the three-dimensional wavelet transform of the original image data that gives rise to some additional sets of radiomics features described, for example, by Guo et al. (Guo X, Liu X, Wang H, et al. Enhanced CT images by the wavelet transform improving diagnostic accuracy of chest nodules. J Digit 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, breaks, and singularities in the original data. At low frequencies (longer time intervals), the wavelets characterize the coarse structure of the data to identify long-term trends. Thus, wavelet analysis enables the extraction of hidden and significant temporal features of the original data while improving the signal-to-noise ratio of imaging studies. The data can be decomposed into multiple (e.g., eight) wavelet decompositions by the discrete wavelet transform by passing the data through a multilevel (e.g., three-level) filter bank. 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. Thus, when a three-level filter bank is used, eight wavelet decompositions corresponding to HHH, HHL, HLH, HLL, LHH, LHL, LLH, and LLL result, where H refers to "high-pass" and L refers to "low-pass". Of course, levels greater than or less than eight can alternatively be used to decompose the data. Such decomposition can be performed using widely available software such as the sliceradiomics software package incorporating the pyradiomics library.When radiomics features are calculated based on wavelet decomposition or transformation of data, this is denoted by a suffix indicating which wavelet decomposition the radiomics feature was calculated based on (e.g., for high-pass, high-pass, high-pass it is HHH). Thus, for example, "Skewness LLL" denotes the radiomics feature "Skewness" calculated based on the LLL wavelet decomposition. In the absence of a suffix, the radiomics feature is calculated based on the original (or raw) data.

[0105] Derivation of a radiomics signature The present invention provides a method for deriving a radiomics signature for characterizing a PVR (e.g., the region 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 including the occurrence of clinical endpoint events for each of the plurality of individuals during a subsequent period after the medical imaging data was collected. In particular, the clinical endpoint preferably indicates cardiovascular health or risk.

[0106] Typically, the method involves performing a case-control study of (human) patients with (clinical endpoint) adverse events versus (human) patients without, over a predetermined or subsequent period, preferably within 5 years, by a clinically required assessment, e.g., by CCTA or other medical imaging techniques, after the imaging data has been collected. Individuals who reach the clinical endpoint are cases (the first group), and individuals who do not reach the clinical endpoint are controls (the second group).

[0107] A plurality of individuals (also referred to herein as patients) can be divided into two independent cohorts of patients who have received (or have had) medical imaging, specifically a training cohort and a validation cohort. Cases are identified based on the occurrence of a specified or predetermined period following the collection of medical imaging data, i.e., clinical endpoint events within a subsequent period. The subsequent period is preferably at least about 5 years, preferably about 5 years, but can be longer or shorter. The clinical endpoint is preferably the primary composite endpoint of major adverse cardiovascular events (MACE), and this primary composite endpoint is a composite of all-cause mortality and non-fatal myocardial infarction (MI) within a specified period (preferably 5 years) following the collection of medical imaging data, or can be defined as the primary composite endpoint of cardiac-specific MACE (cMACE, i.e., cardiac mortality and non-fatal MI) within a specified period (preferably 5 years) following the collection of medical imaging data.

[0108] As used herein, cardiac and non-cardiac mortality can be defined according to ACC / AHA (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 to Develop Cardiovascular Endpoints Data Standards). J Am Coll Cardiol 2015;66(4):403-69). More specifically, cardiac mortality can be defined as any death due to a proximate cardiac cause (e.g., myocardial infarction, low-output heart failure, fatal arrhythmia). Deaths that meet the criteria for sudden cardiac death can also be included in this group. Any death not covered by previous definitions, such as deaths due to malignancy, accident, infection, sepsis, renal failure, suicide, or other non-cardiovascular causes such as stroke or pulmonary embolism, can be classified as non-cardiac.

[0109] Controls are preferably identified as patients with an event-free follow-up within the same specified period, e.g., at least 5 years after CCTA. For example, using an automated algorithm, preferably 1:1 case-control matching is performed to match cases to controls. Cases and controls can be matched for clinical demographics (age, sex, obesity status, etc.), cohort, and / or technical parameters (e.g., tube voltage and CT scanner used) related to imaging data acquisition. Preferably, patients are also matched for other cardiovascular risk factors including hypertension, dyslipidemia, diabetes, and smoking.

[0110] Hypertension can be defined based on the presence of a documented diagnosis or treatment with an antihypertensive regimen according to the 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(25 Pt B):2889-934).

[0111] "Clustering" (or collinearity exclusion) method Next, as schematically illustrated in FIG. 2, a radiomics signature can be generated according to a stepwise approach. First, a plurality of radiomics features are calculated from the medical imaging data of each of a plurality of individuals, for example, as described above. The radiomics features can include the selection or all of the radiomics features as defined in Tables R1 to R7, and each of the radiomics features can be calculated based on the original image data and / or based on one or more wavelet transforms (or wavelet decompositions) of the image data, as described above. Preferably, each of the radiomics features is calculated for the original image data and for the aforementioned eight three-dimensional wavelet decompositions of the image data.

[0112] Unstable features can be removed from the plurality of radiomics features. Z-score transformation can be applied to the features (i.e., representing the values of the radiomics features in terms of the number of standard deviations from the mean value) and the stability analysis based on the z-score. Unstable radiomics features are identified as those having an intraclass correlation coefficient (ICC) in repeated imaging data acquisition (e.g., imaging scans) below a stability threshold. For example, the stability threshold can be at least about 0.9, such as about 0.9, so that all radiomics features having an ICC < 0.9 are excluded (FIG. 3). However, alternatively, other stability thresholds such as 0.85 or 0.95 can be used. The ICC can be calculated for a plurality of repeated scans, such as 2 to 10 scans, particularly 2 or 10 scans. In other words, the stability analysis can be performed on the PVR radiomics features and the unstable radiomics features removed from the plurality of radiomics features. The stability analysis can be performed based on the imaging data of a plurality of individuals or can be performed using other data, such as reference data, for example, the RIDER dataset (a reference image database for evaluating treatment response).

[0113] Preferably, if the plurality of radiomics features include any volume-dependent and orientation-dependent radiomics features, those radiomics features are then excluded (i.e., removed from the plurality of radiomics features). Alternatively, any volume-dependent and orientation-dependent features may be excluded from the start so that such steps are not necessary. In other words, each of the initial plurality of radiomics features may be volume-independent and orientation-independent. Volume-dependent features may include energy and total energy (original and wavelet-computed), and orientation-dependent features may include those derived from wavelet transforms HLL, HLH, LLH, HLL, LHL, LHH (i.e., all except LLL or HHH, or those wavelet transforms that are not exclusively high-pass or low-pass).

[0114] Next, radiomics features that are not significantly associated (e.g., correlated) with clinical endpoints above the significance threshold, such as 5-year MACE or cMACE, can be removed from the plurality of radiomics features. The association of each radiomics feature with the clinical endpoint can be calculated based on data for a plurality of individuals, based on receiver operating characteristic (ROC) analysis, particularly calculation of the area under the curve. The significance threshold is preferably α = 0.05 or less, for example, α can be in the range of 0.001 to 0.05. The significance threshold is preferably approximately α = 0.05. However, the significance threshold may be approximately α = 0.04. Alternatively, the significance threshold may be approximately α = 0.03. Alternatively, the significance threshold may be approximately α = 0.02. Alternatively, the significance threshold may be approximately α = 0.01. Alternatively, the significance threshold may be approximately α = 0.005. Alternatively, the significance threshold may be approximately α = 0.002. Preferably, a radiomics feature must be identified as being significantly associated with the clinical endpoints of both the training cohort and the validation cohort in order to be retained, and anything found to not be significantly associated with the clinical endpoints of one or both cohorts to reduce positive findings due to cohort-specific variation is removed. However, a radiomics feature found to be significantly associated with the clinical endpoint in only one of the cohorts, for example, the training cohort, may alternatively be retained. Alternatively, the assessment of whether each of the radiomics features is significantly associated with the clinical endpoint may be performed based on pooled data for both cohorts, i.e., all of the plurality of individuals. In other words, this method may include assessing the significance of radiomics features using a clinical endpoint and removing features found to not be significantly associated with the clinical endpoint from the plurality of features. The final result should be that any radiomics feature not significantly associated with the clinical endpoint (e.g., as determined or calculated from the data based on analysis of the data) is removed from the plurality of radiomics features.

[0115] Next, the collinearity of the retained radiomics features (i.e., those significantly associated with a clinical endpoint and known as separately significant radiomics features) can be reduced or eliminated by removing pair correlations, i.e., by removing at least one of each pair of collinear radiomics features. The correlation between radiomics features is generally calculated using the measured values of the radiomics features for a plurality of individuals. The removal of pair correlations can be performed in a stepwise manner. Collinear radiomics features can be identified as those that correlate with each other to at least a given correlation threshold. The correlation threshold is preferably applied to both positive and negative correlations and can be expressed, for example, as an absolute value. The pair correlation can be calculated using Spearman's rho coefficient, and the correlation threshold can be at least about |rho| = 0.75, such as about |rho| = 0.75, so that all pair correlations at the level of |rho| ≧ 0.75 are excluded. As will be readily understood in the art, correlation or collinearity is a measure of how closely two radiomics features change together from individual to individual and can be calculated based on the measured radiomics feature values for a plurality of individuals.

[0116] For example, when pairs of collinear radiomics features are identified, one of the two features is preferably excluded from the plurality of features. For example, a radiomics feature that is calculated from the data as not being strongly associated with one of two clinical endpoints may be excluded, and the radiomics feature most strongly associated with the clinical endpoint may be retained, but this is not necessary, and either one can be retained or excluded. For example, the collinearity exclusion step can be performed in a manner that is not monitored without considering the clinical endpoint, and the algorithm can exclude the most redundant features that contribute least to the variation in the study population (e.g., features with smaller variance as measured across multiple individuals). In one example, when pairs of collinear features are identified, the feature with the largest average (e.g., mean) absolute correlation (i.e., the average correlation value (or average absolute value or squared correlation value) with all other radiomics features) is removed. This can be performed stepwise until no collinear radiomics features remain.

[0117] The collinearity exclusion step can be performed using an algorithm or function (see, e.g., the function claret::findCorrelation, R package, Kuhn, M. & Johnson, K. Applied Predictive Modelling. (Springer, 2013)). For example, the function or algorithm can construct a pairwise correlation matrix that includes pairwise correlations between radiomics features. The function can then search the correlation matrix and return a vector of integers corresponding to the columns to be removed to reduce the pairwise correlations. The radiomics features corresponding to these columns can then be removed from the plurality of radiomics features. When determining which column to remove, the algorithm can first identify the pairwise correlations between radiomics features. When two collinear radiomics features are identified, the algorithm can then identify the column corresponding to the feature with the largest average absolute correlation for removal.

[0118] Regardless of how the collinearity exclusion step is performed, the end result is preferably the generation of a reduced plurality of radiomics features where each of the features correlates with each of the other remaining features to an extent less than the correlation threshold. In other words, the method may involve a step of removing radiomics features to remove collinearity among the radiomics features such that none of the remaining radiomics features are collinear with any of the other remaining radiomics features. This may involve the calculation of pairwise correlations between the radiomics features and the removal of at least one of each pair of any identified collinear features.

[0119] Next, a radiomics signature can be constructed based on at least two of the remaining radiomics features that remain after any of the above-described steps (e.g., stability analysis, significance analysis, and / or collinearity exclusion) are performed. For example, a radiomics signature can then be constructed based on at least two of the reduced plurality of non-collinear radiomics features that remain after the collinearity exclusion step. The reduced plurality of features that remain after the collinearity exclusion step are separately known as "original features" herein. However, since each of the excluded radiomics features is strongly correlated with at least one of the original features, a signature in which one or more of the original features are replaced by one of the features that is collinear with the replaced original feature generally functions similarly to a signature calculated based on only the original features. For example, it is possible to exchange one of the original features with one of the features calculated to be collinear with that original feature, and the signature should function similarly. In practice, replacing one or more (or even all) of the original features with alternative features that are collinear with the replaced original features can result in a signature with improved prognostic value, which has in fact been found to be the case in some instances (see, e.g., Tables 8A and 8B). This is because the original features are generally most independently associated with the clinical endpoint, but they are not necessarily the features that function best when combined into a signature.

[0120] Thus, the process of constructing a radiomics signature may involve the construction of "clusters" of radiomics features where each radiomics feature within a cluster is collinear with at least one "original" feature within that cluster (i.e., the features of that cluster remaining after the collinearity exclusion step, e.g., the features within the cluster most strongly independently associated with a clinical endpoint), with each cluster containing one of the original features. The construction of these clusters can be performed instead of the collinearity exclusion step. For example, instead of excluding one of each pair of collinear features, the collinear features may be assigned to the same cluster. Alternatively, the pair exclusion step may be performed as described above, and then, once the original features are identified, the excluded features may be reintroduced by assigning the excluded features to the cluster of original features to which the excluded features are most strongly correlated or collinear.

[0121] However, regardless of how the clusters are constructed, the end result should be that each radiomics feature is assigned to the same cluster as the original radiomics feature(s) with which the radiomics feature is collinear. If a radiomics feature is collinear with two "original" features, it is preferably assigned to the cluster of the original feature with which the radiomics feature is most collinear, but it may be assigned to all clusters of original features with which the radiomics feature is collinear.

[0122] Regardless of whether a radiomics feature itself is independently significantly associated with a clinical endpoint, the cluster can also be extended to include any of the original radiomics features that are collinear with the "original" radiomics feature of that cluster. However, preferably, as described above, any radiomics feature included in the cluster is stable.

[0123] Thus, the "original" radiomics feature within each cluster represents the "partner" radiomics feature for each of the other radiomics features within that cluster, and each of the radiomics features within each cluster is collinear with its "partner" feature. Thus, since the original radiomics feature is completely collinear with itself, it can be regarded as its own "partner" radiomics feature in this sense.

[0124] The first radiomics signature can then be constructed based on at least two (or all) of the reduced number of features (e.g., the "original" features). Alternatively, when clusters are constructed, the first radiomics signature can be constructed from at least two radiomics features each selected from a different cluster. The construction of the radiomics signature involves refining or optimizing the radiomics signature, particularly using data for the "training" cohort. This involves refining or optimizing the contribution of each of the remaining radiomics features to the signature in order to improve the correlation or association of the signature with the clinical endpoint based on the data. For example, the signature can include a weighted sum of the values of each of the radiomics features included in the first signature, and the weight for each of the radiomics features can be progressively optimized or refined. The coefficient multiplied by each of the radiomics features is generally referred to as a beta (β) coefficient, and it is these beta coefficients that can be optimized or refined.

[0125] Preferably, all of the plurality of reduced features are included in the first radiomics signature to be refined, for example, one feature from each cluster, although this is not essential. For example, "top" or "original" features from each cluster may be included (e.g., features most strongly independently associated with the clinical endpoint). Other radiomics features may be included in the first signature to be optimized, for example, two or more radiomics features from any or all of the clusters may be included in the first signature. However, to provide a signature more strongly associated with the clinical endpoint and thus with enhanced diagnostic and prognostic utility, it is preferred to include at least two radiomics features, each from a different cluster. This is because features from different clusters provide complementary information related to PVR. In particular, radiomics 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 first radiomics signature may include at least three radiomics features, each selected from a different cluster. Alternatively, the first radiomics signature may include at least four radiomics features, each selected from a different cluster. Alternatively, the first radiomics signature may include at least five radiomics features, each selected from a different cluster. Alternatively, the first radiomics signature may include at least six radiomics features, each selected from a different cluster. Alternatively, the first radiomics signature may include at least seven radiomics features, each selected from a different cluster. Alternatively, the first radiomics signature may include at least eight radiomics features, each selected from a different cluster. Preferably, the first radiomics signature may include one radiomics feature from each cluster.

[0126] Once the initial radiomics signature is refined, the final radiomics signature can then be constructed based on all of the radiomics features included in the initial signature, or on a subset of those features. For example, the optimized or refined initial signature can be used as the final signature. Alternatively, the final radiomics signature may include only a subset of the radiomics features included in the initial signature. For example, only those parameters having a contribution to the refined initial signature that exceeds a particular threshold may be included in the final signature having an optimized weighting factor (e.g., beta parameter) that exceeds a particular (e.g., predetermined) threshold. In other words, those radiomics features whose coefficients are optimized below a given threshold are removed from the final signature. For example, the top seven features (i.e., those having the greatest contribution to the refined signature) may be included in the final signature.

[0127] If only a subset of the radiomics features is included in the final signature, one option is to re-optimize or refine the signature based on the subset of features. For example, the coefficients or beta parameters can be re-optimized for the subset of radiomics features included in the final signature. Alternatively, coefficients or beta parameters derived from the initial optimization performed on the initial signature can be used in the final signature.

[0128] As described above, a signature can include a weighted sum of calculated values of a plurality of radiomics features. A signature can also include other terms such as addition or subtraction of a constant, or multiplication by a factor. However, typically, a signature is linearly related in some way to the weighted sum of the radiomics feature values.

[0129] The radiomics signature is the term A±Σb i rf i can take the form of, or can include these terms (e.g., the signature can be calculated based on a function that includes these terms), where A is a constant (which can be zero or non - zero), and b i is the weighting coefficient (or beta parameter) for the radiomics feature i, and rf i is the measured value of the radiomics feature i. The beta parameter in the formula is preferably an unnormalized beta obtained by multiplying the Z - score beta parameter by the standard deviation of the radiomics feature, where the standard deviation is preferably the standard deviation of the radiomics feature in the training cohort or the derivation cohort. The constant A is not necessary but can be included to ensure that all resulting values are either positive or negative, preferably positive. Preferably, "±" is "-".

[0130] The first and / or final radiomics signature can be constructed (i.e., optimized or refined) using a machine - learning algorithm. For example, using a machine - learning algorithm, the contribution of each radiomics feature to the signature can be refined or optimized, e.g., by optimizing the beta coefficients. The machine - learning approach can use elastic net / Lasso regression and can utilize leave - one - out cross - validation. Elastic net regression is a regularization regression method that linearly combines the L1 penalty and L2 penalty of the Lasso method and the Ridge method. The L1 (Least Absolute Shrinkage and Selection Operator "Lasso" penalty) imposes a penalty on the absolute value of the coefficients, shrinking irrelevant coefficients to zero and contributing to feature selection, while the L2 ("Ridge" penalty) imposes a penalty on the square of the coefficients, thus limiting the effect of collinearity and reducing overfitting. The optimal penalty coefficient (lambda, λ) can be selected by cross - validation, while alpha is preferably set to α = 1.

[0131] Preferably, the signature is refined using the data of the training cohort and then externally validated using the data of the training cohort.

[0132] "Non-clustering" method The present invention also provides another method for deriving a radiomics signature. In this method, as described above in connection with the "clustering" method, a plurality of radiomics features are calculated from the medical imaging data of each of a plurality of individuals.

[0133] Next, the discrimination values of all radiomics features (preferably, all stable radiomics features) for a clinical endpoint, such as 5-year MACE, can be evaluated using, for example, receiver operating characteristic curve (ROC) analysis, particularly using the calculation of the area under the curve (calculating the C statistic), and radiomics features that are not significantly associated with the clinical endpoint are excluded from the plurality of features. Regarding the clustering method described above, a feature is identified as being significantly associated with a clinical endpoint if the feature is associated with a clinical endpoint that exceeds a specific significance threshold. The significance threshold is preferably α = 0.05 or less, for example, α = 0.001 to 0.05. The significance threshold is preferably approximately α = 0.05. However, the significance threshold may be approximately α = 0.04. Alternatively, the significance threshold may be approximately α = 0.03. Alternatively, the significance threshold may be approximately α = 0.02. Alternatively, the significance threshold may be approximately α = 0.01. Alternatively, the significance threshold may be approximately α = 0.005. Alternatively, the significance threshold may be approximately α = 0.002.

[0134] To correct for multiple comparisons and to reduce the false discovery rate (FDR), a Bonferroni correction can be applied to the significance threshold. The Bonferroni correction can be applied based on the number of principal components that account for a given amount of variation in the study sample based on principal component analysis. For example, the given amount can be approximately 99.5%. In other words, the m value used to correct the α value (by dividing α by m, i.e., α / m) is the number of principal components that account for the given amount of variation. For this reason, principal component analysis of radiomics features can be performed on the data of a plurality of individuals (preferably, including both cohorts if the individuals are divided into a training cohort and a validation cohort).

[0135] Next, a radiomics signature can be constructed based on at least two of the remaining radiomics features, i.e., those identified as being significantly associated with the clinical endpoint, as described above. The radiomics signature can be constructed and optimized in the same manner as described for the above clustering method, except that there is a difference in that it is not necessary to select radiomics features from different clusters.

[0136] For example, all of the radiomics features identified as being significantly associated with the clinical endpoint can be included in the initial or final radiomics signature. Alternatively, only a subset of the radiomics features significantly associated with the clinical endpoint, e.g., at least two, can be included in the initial or final signature. Alternatively, at least three of the radiomics features significantly associated with the clinical endpoint can be included in the initial or final signature. Alternatively, at least four of the radiomics features significantly associated with the clinical endpoint can be included in the initial or final signature. Alternatively, at least five of the radiomics features significantly associated with the clinical endpoint can be included in the initial or final signature. Alternatively, at least six of the radiomics features significantly associated with the clinical endpoint can be included in the initial or final signature. Alternatively, at least seven of the radiomics features significantly associated with the clinical endpoint can be included in the initial or final signature.

[0137] Again, preferably, the signature is refined using the data of the training cohort and then externally validated using the data of the training cohort.

[0138] Elements of either the clustering or non-clustering method can be combined. For example, the clustering (or collinearity exclusion) method can involve the application of a Bonferroni correction.

[0139] For example, a plurality of radiomics features may be input into a machine learning algorithm without performing some (or any) of the foregoing preceding steps. However, typically, before the features are input into the machine learning algorithm, at least the step of excluding unstable radiomics features is performed.

[0140] Generally, in the above-described methods (both clustering and non-clustering), the bivariate association between radiomics features can be evaluated by the non-parametric Spearman's rho (ρ) coefficient, whereas the intraclass correlation coefficient (ICC) can be used to evaluate the within-observer variability in a given number of scans, for example at least 2 scans, for example 10 scans.

[0141] Radiomics signature The PVR radiomics signature of the present invention is calculated based on the measured values of radiomics features obtained from medical imaging data. In particular, the PVR radiomics signature is preferably calculated based on at least two radiomics features.

[0142] To improve the prognostic and diagnostic value of the signature, the signature is preferably calculated based on at least two radiomics features selected from different "clusters" of collinear features, as described above. This reduces redundancy and improves the diversity of information included in the calculation of the signature because features from different clusters are related to different texture aspects of PVR.

[0143] Nine clusters have been identified using the "clustering" method, which corresponds to nine "original" or "partner" features of a reduced plurality of radiomics features remaining after the collinearity exclusion process in the study described in the "Examples" section below. The members of the nine clusters are identified in Table 1.

[0144] Table 1 also identifies the "original features", which are the features within each cluster remaining after the elimination of collinear features in the "clustering" method described above. The collinearity of other features within each cluster with the original features is represented by the |ρ| coefficient (ρ is the non-parametric Spearman's rho (ρ) coefficient).

[0145] To construct the clusters in Table 1, 53 stable volume-independent and direction-independent radiomics features (p < 0.05, p is the probability value or asymptomatic significance, i.e., α = 0.05) associated with clinical endpoints above the threshold were clustered based on the strength of their correlation with the original features in both the training cohort and the validation cohort. When a radiomics feature was found to be collinear with two or more original features (|rho| ≥ 0.75), this radiomics feature was assigned to the cluster of the original feature with the strongest association. 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 demonstrates that the present invention provides a distinct and different prognostic and diagnostic tool compared to calculating FAI PVAT and is thus complementary to FAI PVAT .

Table 1-1

Table 1-2

[0146] As described above, regardless of whether the collinear radiomics features themselves are independently significantly associated with the clinical endpoint, the clusters can be extended to include stable members of multiple original radiomics features that are collinear with the "original" radiomics features of the cluster.

[0147] Members of nine "extended" clusters are presented in Table 2. These extended clusters are equivalent to the original clusters presented in Table 1, but are extended to include any one of 719 stable radiomics features (ICC > 0.9) that are collinear (|rho| ≥ 0.75) with one of the original features from each cluster. Again, if a radiomics feature was found to be collinear (|rho| ≥ 0.75) with more than one original feature, this radiomics feature was assigned to the cluster of original features with which it was found to be most strongly associated.

Table 2-1

Table 2-2

Table 2-3

[0148] The radiomics signature can be constructed from any of the radiomics features (i.e., standard clusters or extended clusters) included in Table 1 or Table 2, provided that the radiomics signature contains at least two radiomics features selected from different clusters. For example, the radiomics signature may contain at least three radiomics features from different clusters. For example, the radiomics signature may contain at least four radiomics features from different clusters. For example, the radiomics signature may contain at least five radiomics features from different clusters. For example, the radiomics signature may contain at least six radiomics features from different clusters. For example, the radiomics signature may contain at least seven radiomics features from different clusters. For example, the radiomics signature may contain at least eight radiomics features from different clusters. For example, the radiomics signature may contain at least nine radiomics features from different clusters.

[0149] Each of at least two (or more) radiomics features selected from different clusters can be selected such that the radiomics feature correlates with the "original" feature of the cluster to which this radiomics feature belongs to at least to the extent of |rho| = 0.8. For example, each of at least two radiomics features from different clusters can correlate with the "original" feature of the cluster to which this radiomics feature belongs to at least to the extent of |rho| = 0.85. For example, each of at least two radiomics features from different clusters can correlate with the "original" feature of the cluster to which this radiomics feature belongs to at least to the extent of |rho| = 0.9. For example, each of at least two radiomics features from different clusters can correlate with the "original" feature of the cluster to which this radiomics feature belongs to at least to the extent of |rho| = 0.95.

[0150] In addition to being calculated based on at least two radiomics features from clusters with different radiomics signatures, the radiomics signature can also be calculated based on additional radiomics features. For example, the radiomics signature can include two or more radiomics features from any given cluster, or can include radiomics features that are not included in any of the clusters.

[0151] Alternatively, the radiomics signature may be calculated based on at least two of the radiomics features that have been found to be independently associated with a clinical endpoint using a "non-clustering" approach. Thus, the radiomics signature may instead be calculated based on at least two radiomics features selected from Table 3. [Table 3-1] [Table 3-2]

[0152] The radiomics signature can be calculated based on at least two radiomics features from Table 3. For example, the radiomics signature can be calculated based on at least three radiomics features from Table 3. For example, the radiomics signature can be calculated based on at least four radiomics features from Table 3. For example, the radiomics signature can be calculated based on at least five radiomics features from Table 3. For example, the radiomics signature can be calculated based on at least six radiomics features from Table 3. For example, the radiomics signature can be calculated based on at least seven radiomics features from Table 3.

[0153] At least two (or more) radiomics features can be selected from those radiomics features in Table 3 that have a Bonferroni-adjusted P-value < 0.04. For example, those radiomics features that have a Bonferroni-adjusted P-value < 0.03. For example, those radiomics features that have a Bonferroni-adjusted P-value < 0.02. For example, those radiomics features that have a Bonferroni-adjusted P-value < 0.01. For example, those radiomics features that have a Bonferroni-adjusted P-value < 0.005. For example, those radiomics features that have a Bonferroni-adjusted P-value < 0.002.

[0154] The signature can be constructed from radiomics features listed in Table 3, excluding the Mean.

[0155] Again, in addition to the radiomics signature being calculated based on at least two radiomics features selected from those presented in Table 3, the radiomics signature can also be calculated based on additional radiomics features. For example, the radiomics signature can include radiomics features not included in Table 3.

[0156] Each of the radiomics signatures of the present invention provides a straightforward means for characterizing PVR using medical imaging data. Since each of the radiomics signatures of the present invention is based on a relatively small number of the total possible radiomics features that can be measured, the signature is easy to calculate and understand, and the physiological significance of the signature can be better understood by clinicians.

[0157] System The method of the present invention can be executed on a system such as a computer system. Accordingly, the present invention also provides a system configured or arranged to execute one or more of the methods of the present invention. For example, the system may include a computer processor configured to execute the method of the present invention, or one or more of the steps of the method. The system may also include a computer-readable memory loaded with executable instructions for performing any of the steps of the method of the present invention.

[0158] In particular, the method of deriving a radiomics signature can be executed on such a system, and accordingly, such a system is provided in accordance with the present invention. For example, the system may be configured to receive and optionally store a data set including values of a plurality of radiomics features of PVR obtained from medical imaging data of each of a plurality of individuals, and information regarding the occurrence during a subsequent period after collection of the medical imaging data of a clinical endpoint indicating the cardiovascular risk of each of the plurality of individuals. The system may be configured to use such a data set to construct (e.g., derive and validate) a radiomics signature according to the method of the present invention.

[0159] Alternatively, the system may be configured to execute a method of characterizing PVR. In particular, the present invention provides a system for characterizing PVR using medical imaging data of a subject. The system may be configured to calculate a value of a radiomics signature of PVR using the medical imaging data. The radiomics signature may be calculated based on measured values of at least two radiomics features of PVR, and the measured values of the at least two radiomics features may be calculated from the medical imaging data.

[0160] The system can also be configured to calculate radiomics features from medical imaging data, as described in more detail above. Thus, the system can be configured to receive and optionally store medical imaging data and process the imaging data to calculate radiomics features.

[0161] Definition of radiomics features The definitions of the radiomics features referred to in this specification are generally well understood within the field of radiomics by merely referring to the names of those radiomics features. However, for ease or reference, the definitions of the features used in this specification are provided in Tables R1 - R7 below. The radiomics features in Tables R1 - R7 are defined according to the radiomics features used by the Pyradiomics package (http: / / pyradiomics.readthedocs.io / en / latest / features.html, see van Griethuysen, J.J.M., Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., Beets-Tan, R.G.H., Fillon-Robin, J.C., Pieper, S., Aerts, H.J.W.L. (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 - R7 comply with the feature definitions described by the Imaging Biomarker Standardisation Initiative (IBSI) which are available in Zwanenburg et al. (2016) (Zwaneburg, A., Leger, S., Valliers, M., and Lock, S. (2016). Image biomarker standardisation initiative - feature definitions. In eprint arXiv:1612.07003[cs.CV]). It should be understood that any definition may be used in accordance with the present invention if the definitions provided below do not exactly comply with the IBSI definitions. Ultimately, the exact mathematical definitions of the radiomics features are not important as minor modifications do not affect the general nature of the images measured by each of the features.Accordingly, minor modifications of the feature quantity (e.g., addition or subtraction of a constant or scaling) and alternative definitions of the feature quantity are intended to be encompassed by the present invention.

[0162] a. Primary statistics These statistics describe the central tendency, variability, uniformity, asymmetry, skewness, and magnitude of the attenuation values in a given region of interest (ROI), ignoring the spatial relationships of individual voxels. Accordingly, those statistics describe the quantitative and qualitative feature quantities of the overall ROI (PVR). A total of 19 feature quantities are calculated for each of the eight wavelet transforms and the original CT image as follows: · Let X be the attenuation or radiation density value (e.g., in HU) of a set of N p voxels included in the region of interest (ROI), · Let P(i) be the number of non-zero bins having a width and equally spaced from 0, and let N g be a primary histogram having N g discrete intensity levels, · Let p(i) be the normalized primary histogram, where

Number

Chemistry

Chemistry

Table 4-1

Table 4-2

[0163] b. Shape-related statistics Shape-related statistics describe the size and shape of a given ROI without considering the attenuation values of the voxels of the given ROI. Since shape-related statistics are independent of gray-level intensity, shape-related statistics are consistent across all wavelet transforms and the original CT images and are therefore calculated only once. They are defined as follows, Let V be the volume of the ROI in mm 3 and Let A be the surface area of the ROI in mm 2 and

Table 5

[0164] c. Gray-level co-occurrence matrix (GLCM) In simple terms, the GLCM describes the number of times a voxel with a given attenuation value i is located adjacent to a voxel of j. Size N g ×N gThe GLCM is constrained by the mask and describes the second-order joint probability density function of the image region defined as P(i,j|δ,θ). The (i,j)-th element of this matrix represents the number of occurrences of the combination of levels i and j in two pixels in the image separated by a distance of δ pixels along the angle θ. The distance δ from the central voxel is defined as the distance following the infinity norm. For δ = 1, this results in 2 neighbors for each of the 13 angles in 3D (26-connectivity), and for δ = 2, it results in 98-connectivity (49 unique angles). To obtain rotationally invariant results, the statistics are calculated in all directions and then averaged to ensure a symmetric GLCM.

[0165]

Chem.

Chem.

Math.

Math.

Math.

Mathematics

Mathematics

Mathematics

Mathematics

Mathematics

Mathematics

Mathematics

[0166] In the case of distance weighting, the GLCM matrix is weighted by the weighting coefficient W and then summed and normalized. The weighting coefficient W is

Mathematics

Table 6-1

Table 6-2

Table 6-3

[0167] d. Gray-Level Size Zone Matrix (GLSZM) The Gray-Level Size Zone (GLSZM) describes the gray-level zones in the ROI, defined as the number of connected voxels sharing the same gray-level intensity. If the distance is 1 according to the infinity norm (26-connected region in 3D, 8-connected region in 2D), the voxels are considered connected. In the gray-level size zone matrix P(i,j), the (i,j)-th element is equal to the number of zones with gray-level i and size j that appear in the image. In contrast to GLCM and GLRLM, GLSZM is independent of rotation, and only one matrix is calculated for all directions in the ROI.

[0168] N g is taken as the number of discrete intensity values in the image, N s is taken as the number of discrete zone sizes in the image, N p is taken as the number of voxels in the image, N z is taken as the number of zones in the ROI, which is

Equation

[0169] e. Gray Level Run Length Matrix (GLRLM) The Gray Level Run Length Matrix (GLRLM) describes the gray level runs defined as the lengths of the number of consecutive pixels having the same gray level value. In the gray level run length matrix P(i,j|θ), the (i,j)-th element describes the number of runs having gray level i and length j occurring in the image (ROI) along the angle θ.

[0170] N g is defined as the number of discrete intensity values in the image, N r is defined as the number of discrete run lengths in the image, N p is defined as the number of voxels in the image, N z (θ) is defined as the number of runs in the image along the angle θ, which is [Math.] equal to, and 1 ≤ Nz ψ(θ) ≤ N p Let P(i, j|θ) be the run-length matrix in an arbitrary direction θ, let p(i, j|θ) be [Number] and be defined as the normalized run-length matrix defined as [Chemistry] is an arbitrarily small positive number (e.g., [Chemistry] 2.2×10 -16 ).

[0171] By default, the feature values are calculated separately for each angle for the GLRLM and then the average of these values is returned. If distance weighting is enabled, the GLRLM is weighted by the distance between neighboring voxels and then summed and normalized. Then, the features are calculated for the resulting matrix. The distance between neighboring voxels is calculated for each angle using the norm specified by "weightingNorm". [Table 8-1] [Table 8-2] [Table 8-3]

[0172] f. Neighboring gray-tone difference matrix (NGTDM) feature The neighboring greylevel difference matrix quantifies the difference between a greylevel and the average greylevel of its neighbors within a distance δ. The sum of the absolute differences for greylevel i is stored in the matrix. X gl Let X be the set of segmented voxels, and x gl (j x ,j y, j z ) ∈ X gl be the greylevel of the voxel at position (j x ,j y, j z ). Then the average greylevel of the neighbors is

Equation

[0173] where W is the number of neighboring voxels also in X gl . Let n i be the number of voxels in X gl with greylevel i, and let Nv,p be the total number of voxels in X gl , equal to Σn i (i.e., the number of voxels with a valid region, at least one neighbor), and N v,p ≤ N p , where N p is the total number of voxels in the ROI. Let p i be the greylevel probability, equal to n i / N v . Let

Equation

Table 9-1

Table 9-2

[0174] g. Gray Level Dependence Matrix (GLDM) The Gray Level Dependence Matrix (GLDM) quantifies the gray level dependence in an image. The gray level dependence is defined as the number of connected voxels within a distance δ that depends on the central voxel. Neighboring voxels with gray level j are considered to depend on the central voxel with gray level i if |i - j| ≤ α. In the gray level dependence matrix P(i,j), the (i,j)-th element describes the number of times a voxel with gray level i appears in the image with a dependence voxel of j in its neighborhood. N g is taken as the number of discrete intensity values in the image, N d is taken as the number of discrete dependence sizes in the image, N z is taken as

Number

Number

Table 10-1

Table 10-2

Example

[0175] Method A two-arm study was conducted designed to explore the diagnostic and prognostic value of coronary PVR radiomic phenotypes on CCTA. The study flowcharts and baseline characteristics of study arms 1 and 2 are presented in Figure 1 and Tables 4A, 4B, and 5. The study protocol was approved by institutional review boards in all regions.

[0176] Arm 1 This was a case-control study comparing the presence or absence of adverse events within 5 years after clinical indication assessment by CCTA. Eligible cases were retrieved from data prospectively collected from two independent cohorts of patients undergoing clinical indication-type CCTA. Of the 4239 individual scans reviewed (2246 in cohort 1 and 1993 in cohort 2), 3912 were of adequate quality and included in further analysis. 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 5 years after CCTA, whereas controls were identified as patients with an event-free follow-up of at least 5 years after CCTA. After review of scan quality and acquisition of relevant demographics, 1:1 matching was performed using an automated algorithm for age, sex, obesity status, cohort, and technical parameters related to CCTA acquisition (tube voltage and CT scanner used). Whenever possible, patients were also matched for other cardiovascular risk factors including hypertension, dyslipidemia, diabetes, and smoking. A subgroup of patients with cardiac-specific MACE (cMACE; cardiac mortality and non-fatal MI) and their matched controls were also examined separately to increase the sensitivity for cardiac-specific high-risk PVR features. [Table 11] [Table 12]

[0177] Arm 2 This was a prospective study (the Ox-IMPACT study, ethical approval provided by the South-Central Oxford C Research Ethics Committee, REC Reference 17 / SC / 0058) recruiting 22 patients presenting with acute myocardial infarction (n = 22 unstable lesions) who were invited to undergo a series of CCTA scans within 96 hours of admission and 6 months later. A control group of 32 patients with known stable CAD (n = 39 stable lesions) and previous percutaneous coronary intervention (PCI) at least 3 months prior to the CCTA scan was also included in this arm. Radiomics phenotypes of coronary PVR were performed around both stable and unstable lesions to identify radiomics signatures of PVR associated with plaque instability and inflammation.

Table 13

[0178] Radiomics features included in the study As outlined in Table 6, a total of 843 PVR radiomics features were measured.

Table 14

[0179] Data collection, definitions, and outcome assessment In Arm 1, outcome data were aggregated through queries of local / national databases not involved in the search of medical records and subsequent image / data analysis. Individual informed consent was waived and approval from the appropriate institutional review board was obtained. Clinical data and demographics were recorded prospectively in the electronic medical record at the first clinical encounter and manually extracted for the current study. Hypertension was defined based on the presence of documented diagnosis or treatment with an antihypertensive regimen according to the 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 were 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(25 Pt B):2889-934).The exact determination of the cause of death was performed locally by the study principal investigator at each facility through review of the charts, examination of the death certificates, and / or telephone follow-up and / or verification involving the family. Cardiac and non-cardiac mortality were defined according to the recommendations of the ACC / AHA and the Academic Research Consortium, as described above. More specifically, cardiac mortality was defined as any death due to a proximate cardiac cause (e.g., myocardial infarction, low-output heart failure, fatal arrhythmia). Deaths meeting the criteria for sudden cardiac death were also included in this group. Any death not covered by the previous definition, such as deaths due to malignancy, accident, infection, sepsis, renal failure, suicide, or other non-cardiovascular causes such as stroke or pulmonary embolism, was classified as non-cardiac. Deaths for which information on the cause of death could not be reliably collected were classified as "death of unknown cause" at the discretion of the study principal investigator at the local facility. Non-fatal myocardial infarction events (ST-segment elevation or non-ST-segment elevation) during follow-up were also retrieved through searches of the electronic health records.

[0180] Coronary Computed Tomography Angiography (CCTA) Acquisition Protocol Cohort 1: Most (87.1%) of the CCTA scans were performed on a 256-slice Brilliance iCT scanner (Philips Medical Systems, Best, The Netherlands), and the remainder were 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 per minute, if the heart rate remained above 60 beats per minute after being positioned on the CT table, intravenous metoprolol 5 mg (increased in 5 mg increments up to a maximum dose of 30 mg) or intravenous diltiazem 5 mg (increased in 5 mg increments up to a maximum dose of 20 mg) was administered. Patients were also dosed sublingually with 0.3 mg of nitroglycerin immediately before CCTA, and an iodinated contrast agent (Omnipaque 350, General Electric, Milwaukee, USA) was administered at a flow rate of 5–6 ml / s.

[0181] Cohort 2: CCTA scans were performed using a 2×64-slice scanner (Definition Flash, Siemens Healthcare, Forchheim, Germany) (79.2%), and the remainder using either a 64-slice (Siemens Sensation 64, Siemens Healthcare, Forchheim, Germany) (18.1%) or a 2×128-slice scanner (2.7%) (Somatom Definition Flash, Siemens Healthcare, Forchheim, Germany). When the heart rate >60 beats per minute, an oral dose of 100 mg atenolol was administered 1 hour before CT, and an additional 5 mg metoprolol up to a maximum dose of 30 mg was administered intravenously if the heart rate remained >60 beats per minute after the patient was positioned on the CT table. Patients were also dosed sublingually with 0.8 mg nitroglycerin immediately before CCTA, and an iodinated contrast agent (Omnipaque 350, Schering AG, Berlin, Germany) was administered at a flow rate of 5–6 ml / s.

[0182] Study arm 2: Participants in study arm 2 underwent CCTA using a 64-slice scanner (General Electric, LightSpeed Ultra, General Electric, Milwaukee, WI, USA). Heart rate was optimized using intravenous injection of a beta blocker, and sublingual glyceryl-trinitrate (800 μg) was also administered to achieve maximal coronary artery dilation. 95 ml of an iodinated contrast agent (Niopam 370, BRACCO UK Ltd) was injected intravenously at a flow rate of 6 mL / second (120 kVp tube energy, 0.625 mm axial slice thickness, 0.35 second rotation time, 40 mm detector coverage), after which CCTA was performed. Prospective image acquisition was used with ECG gating at 75% of the cardiac cycle (with 100 msec padding for optimal imaging of the right coronary artery if required).

[0183] Processing and analysis of CCTA scans All images were first locally anonymized and then transferred to the Core Lab (Academic Cardiovascular CT Unit, University of Oxford, United Kingdom) for analysis by blinded clinical trial responsible physicians for the population demographics and outcomes of the cohort on a dedicated workstation (Aquarius Workstation® V.4.4.13, TeraRecon Inc., Foster City, CA, USA). All scans were first reviewed based on scan quality and the presence of artifacts that would interfere with reliable qualitative and quantitative assessment. Low-quality scans were reviewed by at least two independent clinical trial responsible physicians before being excluded from subsequent analysis. Mild, moderate, and severe coronary stenosis were defined as lumen stenosis of 25–49%, 50–69%, and ≥70%, respectively (as previously described in Cury RC, Abbara S, Achenbach S, et al. CAD-RADS® 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). Endorsed by the American College of Cardiology. J Cardiovasc Comput Tomogr 2016;10(4):269-81).Obstructive coronary artery disease (CAD) was defined as the presence of ≥1 coronary artery lesion causing ≥50% luminal stenosis, but the extent of CAD was evaluated by the Duke Prognostic CAD Index (e.g., 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 1 of the following features on CCTA, namely, a) spotty calcification, b) low-attenuation plaque, c) positive remodeling, and d) napkin-ring sign. (As previously described 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 evaluated by measuring the total volume of epicardial adipose tissue (EAT) in a semi-automated manner by tracing the contour of the pericardium from the level of the pulmonary artery bifurcation to the most caudal cardiac apex. FAI. PVAT was defined as previously described in Antonopoulos AS, Sanna F, Sabharwal N, et al. “Detecting human coronary inflammation by imaging perivascular fat.” Sci Transl Med 2017;9(398).

[0184] Coronary PVR radiomics feature extraction Calculation of radiomics features in coronary PVR was performed on all selected CCTA scans using 3D Slicer (v.4.9.0-2017-12-18 r26813, available from http: / / www.slicer.org). To avoid the problem of collinearity between different coronary arteries, the analysis in arm 1 was restricted to the proximal and mid-right coronary artery (RCA) (segments 1 and 2 according to the American Heart Association's anatomical classification). Coronary PVR was defined as all voxels within the Hounsfield unit range of -190 to -30 Hounsfield units (HU) located within a radial distance equal to the diameter of the adjacent vessel from the outer vessel wall. Segmentation of coronary PVR was performed by placing three-dimensional spheres with a diameter equal to three times the diameter of the coronary artery along the centerline of the vessel on consecutive slices. Subsequently, the segmented PVR was extracted and used to calculate radiomics features using the SlicerRadiomics extension of 3D Slicer that incorporates the Pyradiomics library into 3D Slicer. Shape-related and first-order radiomics features were calculated using the original HU values of the segmented PVR. For the calculation of texture features (gray-level co-occurrence matrix [GLCM], gray-level dependence matrix [GLDM], gray-level run length matrix [GLRLM], gray-level size zone matrix [GLSZM], and neighborhood gray-tone difference matrix [NGTDM], Fig. 1A, Tables R1-R7), the PVR voxels were discretized into 16 bins of equal width (10 HU width) to enable sufficient resolution to detect biologically significant spatial variations in PVR attenuation while reducing noise.To enforce symmetric, rotation-invariant results, texture statistics (such as GLCM) were calculated in all four directions and then averaged (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).

[0185] Wavelet transform: Also, for the three-dimensional wavelet transform of the original image, first-order statistics and texture-based statistics were calculated, resulting in eight additional sets of radiomic features (Figure 1A).

[0186] Statistical analysis In arm 1, case-control matching was performed using the automated algorithm provided by the ccmatch command in Stata. In the final study population, clinical demographics were presented as the mean ± standard deviation for continuous variables and as percentages for categorical variables. Continuous variables between the two groups were compared by Student's t-test, while categorical variables were compared using Pearson's chi-square test.

[0187] Principal component and unsupervised clustering: In both study arms, all 843 calculated PVR radiomic features were included in a principal component analysis to identify the principal components that describe most of the phenotypic variation in the study population. Then, hierarchical clustering of the observations was performed using the first three components in arm 1 (PC1, PC2, PC3) and arm 2 (PC1’, PC2’, and PC3’) (using Ward's method and Minkowski distance). The frequencies of MACE (arm 1) and unstable plaques (arm 2) between distinct clusters were compared by chi-square test.

[0188] Improvement in Feature Selection and Discrimination: First, in a receiver operating characteristic (ROC) analysis, the discrimination values of all radiomics features for 5-year MACE were tested. This analysis was performed on pooled data for both cohorts. To correct for multiple comparisons, a Bonferroni correction was applied to the number of principal components accounting for 99.5% of the variability in our study sample, following an approach based on genome-wide association studies (GWAS) that has previously been used in the field of radiomics image analysis (see, for example, 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) 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). The bivariate association between radiomics features was evaluated by the non-parametric Spearman's rho (ρ) coefficient, whereas intra-observer variability was evaluated by the intra-class correlation coefficient (ICC) over 10 scans.

[0189] To form the radiomics signature of high-risk PVR, a multi-step approach was followed. Z-score transformation was applied to all features, and unstable radiomics features with low ICC in iterative analysis (<0.9) were excluded. To minimize false-positive findings driven by cohort-specific variations, features significantly associated with MACE in both cohorts (at the α = 0.05 level) were then selected. Next, collinearity was reduced by stepwise removal of paired comparisons using appropriate functions of the caret package in R. Then, a machine learning approach was applied in cohort 1 using elastic net regression and leave-one-out internal cross-validation. The optimal penalty coefficient (lambda, λ) was selected by cross-validation, while alpha was set to α = 1. Next, the best performance models for both MACE and cMACE were externally validated in cohort 2, and discrimination was evaluated by calculating the area under the curve (AUC). Then, the top variables of the best performance model were combined into a unified score / signature by multiplying the unadjusted beta coefficient of each variable of the model and calculating the sum. The unified signature was then added in a logistic regression model consisting of the following four blocks: i) age, gender, hypertension, dyslipidemia, smoking, and diabetes (model 1), ii) model 1 + CT-derived measurements, where CT-derived measurements include the Duke Prognostic CAD index, presence of high-risk plaque features, EAT volume and presence of coronary artery calcium (model 2), iii) model 2 + FAI PVAT (model 3), and iv) model 3 + PVR texture signature. The prognostic values of the nested models of MACE and cMACE were compared by the respective C-statistics (area under the curve, AUC) of the models. FAI PVATThe interaction between and the PVR texture signature is presented graphically using bivariate contour plots derived from multiple previous logistic regression models. The comparison of selected radiomics features between unstable and stable plaques in arm 2 was performed using the non-parametric Mann-Whitney test. Statistical analyses were performed in the R environment (packages: caret, hclust), as well as in Stata v14.0 (Stata Corp Inc., College Station, Texas). Unless otherwise specified, all tests were two-sided and α was set at 0.05.

[0190] Results Study cohort demographics Of the 3912 individual scans reviewed in arm 1, a total of 386 scans were included in the study, corresponding to 193 patients with 5-year MACE and 193 matched controls. A selected subgroup of 98 patients with cardiac-specific MACE (cMACE) and their matched controls was also identified and analyzed in a subgroup type analysis. The demographics and baseline characteristics of the cohort in study arm 1 are outlined in Tables 4A and 4B. Cases and controls did not differ significantly in terms of baseline demographics, but as expected, cases (MACE or cMACE) were more likely to have coronary artery disease (CAD) as assessed by the degree of coronary stenosis and the Duke Prognostic CAD Index compared to controls. Study arm 2 included 22 patients with unstable lesions, as well as 32 age- and sex-matched controls with a total of 39 stable lesions (Table 5).

[0191] PVR radiomics: component analysis, and association with adverse events In addition to the individual features of 18 first-order statistics, 23 GLCM, 14 GLDM, 16 GLRLM, 16 GLSZ, and 5 NGTDM, a total of 103 radiomics features corresponding to 15 shape-related features were calculated from the original images (Table 6, Figure 4). All features except the shape-related features were also calculated for each of the 8 wavelet transforms of the segmented regions, resulting in 843 radiomics features. Principal component analysis of all radiomics features revealed three principal components that accounted for 61.98% of the variance observed in both cohorts of Arm 1 (Figure 5A), while 92 components accounted for 99.5% of the variance in the study population (scree plot presented in Figure 5B). Of the 42 components with eigenvalue >1, a total of 6 components were significant predictors of MACE (Figure 5C), while 4 components remained independently associated with MACE in a multiple regression model that included all 5 components adjusted for age, gender, cardiovascular risk factors, and indices from CCTA of CAD severity (Figure 5D). These components were differentially associated with baseline clinical demographics and characteristics (Figure 5E) and, in some cases, reflected features of PVR that described distinct biological phenotypes. Notably, inclusion of 4 PVR components in a model that included age, gender, cardiovascular risk factors, and risk features from CCTA significantly improved discrimination of 5-year MACE (Figure 5F), suggesting that the radiomics phenotype of PVR holds a strong and incremental prognostic value.

[0192] Unsupervised clustering based on coronary PVR phenotype Unsupervised (hierarchical) clustering of the pooled arm 1 study population using the first three principal components (PC1, PC2, and PC3) of coronary PVR radiomics identified three distinct clusters with significantly different risks of 5-year MACE (46.5% vs. 45.8% vs. 65.8% MACE, P = 0.009). Similarly, in arm 2, hierarchical clustering of the identified coronary lesions identified two distinct clusters with significantly different prevalences of unstable plaque (58.8% vs. 25%, P = 0.01). These findings suggest the presence of an increased risk of cardiovascular disease and distinct radiomics signatures in PVR associated with both the local presence of coronary inflammation and unstable lesions.

[0193] Identification of specific high-risk PVR radiomics features Principal component analysis and unsupervised clustering demonstrated the concept that the radiomics phenotype of PVR can be associated with both the local presence of coronary inflammation / disease and adverse outcomes. However, they were unable to identify specific high-risk radiomics features that could be reproducibly measured in independent cohorts. The principal components are specific to the dataset from which they are derived and are not readily applicable to independent datasets and cohorts. Therefore, established FAI PVATFurther work was done to identify specific features that can complement the diagnostic and prognostic information of the markers. In the ROC analysis for the discrimination of the primary endpoint of MACE, a total of 198 features were found to be significant at the 0.05 level using the pooled data for both cohorts. To correct for multiple comparisons and reduce the false discovery rate (FDR), 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 as significant discriminators of MACE, as outlined in the Manhattan plot (Figure 6) and Table 3. Of these 46 radiomics features, 28 were first-order statistics, 10 were GLRLM, 5 were GLCM, 1 was GLDM, 1 was GLSZ, and 1 was NGTDM, while half were derived from the original images and the other half were derived from wavelet transforms (17 LLLs, 2 HHHs, 3 HHLs, and 1 LHL).

[0194] Formation of a high-risk PVR radiomics signature FAI PVATIndividual radiomic features such as have been associated with an increased cardiovascular risk, but whether combinations of different radiomic features (a "signature") can provide a more powerful way to characterize the adverse profile of coronary PVR has been previously unknown. To examine this hypothesis and develop a radiomic signature that is both prognostic and reproducible, a rigorous stepwise approach was applied (Figure 2). First, of the 843 computed coronary PVR radiomic features, 143 were excluded after stability analysis (ICC < 0.9, two measurements) (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 potential false-positive findings due to cohort-specific variation. Subsequently, collinearity was reduced by stepwise exclusion of pair correlations at the level of |rho| ≧ 0.75 (Spearman's 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 prediction of MACE showed average discrimination (Figure 2B), and a similar model for prediction of cardiac-specific MACE (Figure 2C) showed very good discrimination values in both cohort 1 (derivation) and cohort 2 (validation cohort). Then, as shown in Table 7, using the top six components of this model, a perivascular texture index (PTI) was defined using the coefficients derived from the model. Thus, Table 7 describes a specific example of the radiomic signature of the present invention. [Table 15]

[0195] In Table 7, the z-score beta was standardized beta (b z ) was converted by multiplying by the standard deviation of each variable in the derivation cohort (cohort 1). A constant of 90 was added post hoc to ensure positive values based on the range of values observed in all cohorts analyzed. i )

[0196] Incremental value of the PVR radiomics phenotype beyond the current state of the art As shown in FIGS. 3A and 3B, FAI beyond the current risk biomarkers used in CCTA-based risk stratification PVAT To evaluate the incremental value of the radiomics phenotype, a set of nested models was constructed. Model 1 consisted of demographics and conventional risk factors, which showed inadequate discrimination for both MACE and cMACE, partly due to the matching process for the selection of the control group. The addition of risk features from CCTA (including the Duke Prognostic CAD Index, presence of coronary artery calcium, high-risk plaque features, and epicardial adipose tissue (EAT) volume) significantly improved the discrimination for both MACE and cMACE, while the inclusion of PVAT FAI further increased the discrimination for both endpoints. Despite this marked improvement, the addition of PTI (as defined in Table 7) was associated with a further significant improvement in the discriminative value of the model for both endpoints (FIGS. 3A and 3B), suggesting not only an independent value but also an incremental value of the PVR texture phenotype in risk prediction. The interaction between FAI PVAT and PTI for the prediction of MACE / cMACE is graphically presented in FIG. 3C, demonstrating that for a given FAI PVAT value, there is an increasing risk with higher levels of PTI and vice versa.

[0197] Validation of an alternative radiomics signature (PTI) of the present invention The data presented in FIGS. 3A and 3B demonstrate that the radiomics signature (PTI) as defined in Table 7 provides a significant improvement in the discriminative value of the model for both endpoints (MACE and cMACE). To validate the usefulness of alternative radiomics signatures of the present invention, including different selections of radiomics features, a series of models including several different radiomics signatures were tested against the current state-of-the-art model.

[0198] In Table 8, the improvement in model performance is presented for 196 patients (98 with cardiac MACE and 98 matched controls). Each step corresponds to including one selected radiomics feature on top of the current state-of-the-art model and the radiomics features of the previous cluster. The current state-of-the-art model includes age, gender, hypertension, dyslipidemia, smoking and diabetes, Duke Prognostic Coronary Artery Disease index, presence of high-risk plaque features, epicardial adipose tissue volume, and presence of coronary artery calcium.

[0199] In each of Examples 1 - 4, the current state-of-the-art model was progressively complemented by a radiomics signature that included progressively more radiomics features from different clusters. First, the state-of-the-art model was complemented by a radiomics signature calculated based on radiomics features selected from Cluster 1 (first row in Tables 8A and 8B: “+Cluster 1”). Next, the state-of-the-art model was complemented by a radiomics signature calculated based on two radiomics features selected from Clusters 1 and 2 (first row in Tables 8A and 8B: “+Cluster 2”). Thus, each progressive row in Tables 8A and 8B corresponds to including one selected radiomics feature on top of the current state-of-the-art model and the radiomics features of the previous cluster. Nagelkerke's pseudo-R 2 provides a measure of discrimination of the model for cMACE.

[0200] From Tables 8A and 8B, it is clearly seen that all signatures of the present invention calculated based on different selections of radiomics features from the identified clusters provide an improved prediction of cardiac-specific MACE. Thus, the data presented in Table 8 demonstrate that the radiomics signature (PTI) of the present invention provides an improved prediction of cardiovascular risk compared to previously used models, regardless of which features are selected from each of the identified clusters or exactly how many are selected.

Table 16

Table 17

[0201] PVR radiomics phenotype for detecting unstable lesions When calculated around predefined coronary lesions in study arm 2, all but one of the identified radiomics features used to define PTI were significantly modified in the presence of unstable culprit lesions (scanned within 96 hours of ACS onset) compared to stable treated lesions (scanned >3 months after PCI) (Figs. 8A - G). More importantly, PTI (defined using the formula in Table 7) outperformed all six individual radiomics features in differentiating unstable morphologically stable coronary lesions (AUC: 0.76; 95% confidence interval: 0.64 - 0.88) (Fig. 8G). These findings suggest the value of the combination of PTI and FAI with the perivascular fat phenotype for both future risk prediction and detection of unstable coronary plaques, and clarify the close link between the perivascular fat phenotype and vascular inflammation, plaque instability, and adverse clinical outcomes. PVAT and clarify the close link between the perivascular fat phenotype and vascular inflammation, plaque instability, and adverse clinical outcomes.

[0202] Summary of Findings Using a machine learning approach, the inventors added incremental value beyond traditional risk factors and established CCTA risk classification tools for predicting future adverse events and assessing cardiovascular health and risk, and discovered and validated a coronary PVR radiomics signature that further detects the presence of local plaque inflammation and unstable coronary lesions. The inventors demonstrated that a PVR radiomics signature based on two or more radiomics features of PVR provides a tool for predicting future adverse events in patients, diagnosing coronary artery disease or coronary heart disease, and identifying unstable coronary lesions.

[0203] The PVR signature of the present invention describes a high-risk PVR phenotype that links the PVR signature to future event risks and provides incremental prognostic information beyond current CCTA-based tools. The signature of the present invention can also distinguish unstable coronary artery lesions from stable coronary artery lesions based on the radiomics signature of perivascular fat. In summary, the findings presented herein demonstrate that the PVR radiomics phenotype using the radiomics signature of the present invention can identify both high-risk patients (when measured in a standardized manner around coronary blood vessels) and high-risk lesions (when applied around specific coronary segments or lesions), which has important implications for risk prediction based on the latest CCTA.

[0204] Surprisingly, the radiomics signature need not be constructed from the radiomics features most strongly and independently associated with future adverse events. Instead, it is actually advantageous to include the selection of radiomics features from different collinear "clusters" of radiomics features, rather than simply including the radiomics features most associated with adverse events individually.

[0205] A particularly attractive aspect of the present invention is that this aspect can be implemented on previously collected historical medical imaging data. The signature of the present invention may 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 the need to perform additional scans. Therefore, the method of the present invention need not include the step of collecting medical imaging data and can be implemented based on the retrospective analysis of existing medical imaging data.

[0206] Selected aspects of the present invention The following numbered clauses disclose various aspects of the present invention. Clause 1. A method for characterizing a perivascular region using medical imaging data of a subject, comprising calculating a value of a radiomics signature of the perivascular region using the medical imaging data, A method in which a radiomics signature is calculated based on measured values of at least two radiomics features in a perivascular region, and the measured values of the at least two radiomics features are calculated from medical imaging data.

[0207] Clause 2. At least two radiomics features are selected from the radiomics features of clusters 1 to 9, and each of the at least two radiomics features is selected from a different cluster. Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, Autocorrelation, Sum Average, Joint Average, and High Gray Level Zone Emphasis. 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, Dependence Entropy LLL, Dependence Entropy, Zone Entropy LLL, Run Entropy LLL, and Mean LLL. Cluster 4 consists of Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, Short Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Low Gray Level Emphasis, Small Dependence Low Gray Level Emphasis, Gray Level Variance LLL(GLSZM), Gray Level Variance(GLDM), Variance, Gray Level Variance(GLDM), Difference Variance LLL, Gray Level Variance LLL(GLRLM), Variance LLL, Gray Level Variance LLL(GLDM), Sum of Squares, Contrast LLL, Mean Absolute Deviation, Interquartile Range, Robust Mean Absolute Deviation, Long Run Low Gray Level Emphasis, Difference Variance, Gray Level Variance(GLSZM), Inverse Difference Moment Normalized, Mean Absolute Deviation LLL, Sum of Squares LLL, and Contrast, Cluster 5 consists of Zone Entropy, Gray Level Non Uniformity Normalized (GLRLM), Gray Level Non Uniformity Normalized LLL (GLRLM), Sum Entropy, Joint Energy, Entropy, Gray Level Non Uniformity Normalized (GLDM), Joint Energy, Gray Level Non Uniformity Normalized LLL (GLDM), Uniformity LLL, Sum Entropy LLL, and Uniformity (Zone Entropy, Gray Level Non Uniformity Normalized(GLRLM), Gray Level Non Uniformity Normalized LLL(GLRLM), Sum Entropy, Joint Energy, Entropy, Gray Level Non Uniformity Normalized(GLDM), Joint Energy, Gray Level Non Uniformity Normalized LLL(GLDM), Uniformity LLL, Sum Entropy LLL, and Uniformity), Cluster 6 consists of Zone Entropy HHH, Size Zone Non Uniformity Normalized HHH, and Small Area Emphasis HHH (Zone Entropy HHH, Size Zone Non Uniformity Normalized HHH, and Small Area Emphasis HHH), Cluster 7 consists of Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, Coarseness LLH, Coarseness HHH, Coarseness HLH, Coarseness HHL, and Coarseness LHH (Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, Coarseness LLH, Coarseness HHH, Coarseness HLH, Coarseness HHL, and Coarseness LHH), Cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level Variance LLL (GLSZM), Sum of Squares, Robust Mean Absolute Deviation, Sum Entropy, Interquartile Range, Cluster Prominence LLL, Entropy LLL, 10th Percentile LLL, 10th Percentile (Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance(GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance(GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level Variance LLL (GLSZM), Sum of Squares, Robust Mean Absolute Deviation, Sum Entropy, Interquartile Range, Cluster Prominence LLL, Entropy LLL, 10th Percentile LLL, 10th Percentile), Cluster 9 has size zone non-uniformity LLL, dependence non-uniformity HLL, gray level non-uniformity HLL (GLSZM), gray level non-uniformity (GLSZM), run length non-uniformity HHL, run length non-uniformity LHL, dependence non-uniformity LHL, dependence non-uniformity, run length non-uniformity HLH, busyness, run length non-uniformity LLH, dependence non-uniformity LLH, dependence non-uniformity LLL, size zone non-uniformity, energy HLL, run length non-uniformity LHH, size zone non-uniformity HLL, gray level non-uniformity LLH (GLSZM), gray level non-uniformity LHL (GLSZM), gray level non-uniformity LLL (GLSZM), run length non-uniformity HLL, gray level nonThe method according to claim 1, comprising Uniformity HLH (GLSZM), Gray Level Non Uniformity HHL (GLSZM), Run Length Non Uniformity, and Run Length Non Uniformity HHH.

[0208] Clause 3. Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, Autocorrelation, Sum Average, and Joint Average. Cluster 2 consists of Skewness, Skewness LLL, Kurtosis, 90th Percentile, 90th Percentile LLL, Median LLL, and Kurtosis LLL. Cluster 3 consists of Run Entropy, Dependence Entropy LLL, Dependence Entropy, and Zone Entropy LLL. Cluster 4 consists of Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, Short Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Low Gray Level Emphasis, Small Dependence Low Gray Level Emphasis, Gray Level Variance LLL(GLSZM), Gray Level Variance(GLDM), Variance, Gray Level Variance(GLDM), and Difference Variance LLL, Cluster 5 consists of Zone Entropy, Gray Level Non Uniformity Normalized(GLRLM), and Gray Level Non Uniformity Normalized LLL(GLRLM), Cluster 6 consists of Zone Entropy HHH and Size Zone Non Uniformity Normalized HHH, Cluster 7 consists of Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, Coarseness LLH, and Coarseness HHH, Cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level 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 (Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance(GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance(GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level 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), Cluster 9 consists of Size Zone Non Uniformity LLL, Dependence Non Uniformity HLL, Gray Level Non Uniformity HLL (GLSZM), Gray Level Non Uniformity (GLSZM), Run Length Non Uniformity HHL, Run Length Non Uniformity LHL, Dependence Non Uniformity LHL, Dependence Non Uniformity, Run Length Non Uniformity HLH, Busyness, Run Length Non Uniformity LLH, Dependence Non Uniformity LLH, Dependence Non Uniformity LLL, Size Zone Non Uniformity, Energy HLL, Run Length Non Uniformity LHH, Size Zone Non Uniformity HLL, Gray Level Non Uniformity LLH (GLSZM), and Gray Level Non Uniformity LHL (GLSZM) (as described in Article 2).

[0209] Article 4. Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, Autocorrelation, Sum Average, and Joint Average, Cluster 2 consists of Skewness, Skewness LLL, Kurtosis, and 90th Percentile, Cluster 3 consists of Run Entropy, Dependence Entropy LLL, and Dependence Entropy, Cluster 4 consists of Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, Short Run Low Gray Level Emphasis, Low Gray Level Run Emphasis, Low Gray Level Emphasis, and Small Dependence Low Gray Level Emphasis, Cluster 5 consists of Zone Entropy, and Gray Level Non Uniformity Normalized (GLRLM), Cluster 6 consists of Zone Entropy HHH, Cluster 7 consists of Strength, Coarseness HLL, Coarseness, Coarseness LHL, Coarseness LLL, and Coarseness LLH, Cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL(GLDM), Variance LLL, Gray Level Variance LLL(GLRLM), Gray Level Variance(GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance(GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level 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, Cluster 9 consists of Size Zone Non Uniformity LLL, Dependence Non Uniformity HLL, Gray Level Non Uniformity HLL (GLSZM), Gray Level Non Uniformity (GLSZM), Run Length Non Uniformity HHL, Run Length Non Uniformity LHL, Dependence Non Uniformity LHL, Dependence Non Uniformity, Run Length Non Uniformity HLH, Busyness, Run Length Non Uniformity LLH, Dependence Non Uniformity LLH, Dependence Non Uniformity LLL, and Size Zone Non Uniformity, and is the method according to Clause 2.

[0210] Clause 5. Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, High Gray Level Run Emphasis, and Autocorrelation, Cluster 2 consists of Skewness, and Skewness LLL, Cluster 3 consists of Run Entropy, and Dependence Entropy LLL, Cluster 4 consists of Small Area Low Gray Level Emphasis, and Low Gray Level Zone Emphasis, Cluster 5 consists of Zone Entropy, Cluster 6 consists of Zone Entropy HHH, Cluster 7 consists of Strength, Cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level 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 (Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level 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), Cluster 9 consists of Size Zone Non Uniformity LLL, Dependence Non Uniformity HLL, Gray Level Non Uniformity HLL (GLSZM), Gray Level Non Uniformity (GLSZM), Run Length Non Uniformity HHL, and Run Length Non Uniformity LHL, according to the method described in Clause 2.

[0211] Clause 6. Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, and High Gray Level Run Emphasis, Cluster 2 consists of Skewness, and Skewness LLL, Cluster 3 consists of Run Entropy, Cluster 4 consists of Small Area Low Gray Level Emphasis, and Low Gray Level Zone Emphasis, Cluster 5 consists of Zone Entropy, Cluster 6 consists of Zone Entropy HHH, Cluster 7 consists of Strength, Cluster 8 consists of Cluster Tendency LLL, Cluster Tendency, Sum of Squares LLL, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Gray Level Variance (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance (GLDM), Variance, Mean Absolute Deviation, Cluster Prominence, Sum Entropy LLL, Interquartile Range LLL, Gray Level Variance LLL (GLSZM), Sum of Squares, Robust Mean Absolute Deviation, and Sum Entropy, Cluster 9 consists of Size Zone Uniformity LLL (Size Zone Uniformity LLL), the method according to clause 2.

[0212] Clause 7. Cluster 1 consists of Short Run High Gray Level Emphasis, High Gray Level Emphasis, and High Gray Level Run Emphasis, Cluster 2 consists of skewness, skewness LLL, kurtosis, 90th percentile, median LLL, kurtosis LLL, and median (Skewness, Skewness LLL, Kurtosis, 90th Percentile, Median LLL, Kurtosis LLL, and Median), Cluster 3 consists of run entropy, run entropy LLL, and mean LLL (Run Entropy, Run Entropy LLL, and Mean LLL), Cluster 4 consists of small area low gray level emphasis, low gray level zone emphasis, and gray level variance (GLSZM) (Small Area Low Gray Level Emphasis, Low Gray Level Zone Emphasis, and Gray Level Variance (GLSZM)), Cluster 5 consists of zone entropy, gray level non-uniformity normalized (GLRLM), and uniformity (Zone Entropy, Gray Level Non Uniformity Normalized (GLRLM), and Uniformity), Cluster 6 consists of zone entropy HHH (Zone Entropy HHH), Cluster 7 consists of strength (Strength), Cluster 8 is characterized by Cluster Tendency LLL, Cluster Tendency, Mean Absolute Deviation LLL, Gray Level Variance LLL (GLDM), Variance LLL, Gray Level Variance LLL (GLRLM), Robust Mean Absolute Deviation LLL, Gray Level Variance LLL (GLSZM), Interquartile Range LLL, Sum Entropy LLL, Gray Level Variance (GLRLM), Variance, Gray Level Variance (GLDM), Mean Absolute Deviation, Sum Entropy, Robust Mean Absolute Deviation, Interquartile Range, Entropy LLL, 10th Percentile LLL, 10th Percentile, Entropy, Uniformity LLL, Gray Level Non Uniformity Normalized LLL (GLDM), Gray Level Non Uniformity Normalized LLL (GLRLM), Root Mean Squared, Gray Level Non Uniformity Normalized (GLDM), Root Mean Squared LLL, and Longconsisting of Run Low Gray Level Emphasis), The method according to clause 2, wherein cluster 9 consists of Size Zone Non Uniformity LLL, Busyness, and Size Zone Non Uniformity.

[0213] Clause 8. At least two radiomics features are the median LLL, mean LLL, median, root mean squared LLL, mean, kurtosis, root mean squared, run entropy LLL (GLRLM), uniformity, 90th percentile, gray level non-uniformity normalized (GLRLM), uniformity LLL, skewness, gray level non-uniformity normalized LLL (GLRLM), 10th percentile LLL, skewness LLL, 10th percentile, entropy, interquartile range LLL, robust mean absolute deviation LLL, run entropy (GLRLM), interquartile range, sum entropy (GLCM), gray levelThe method according to claim 1, which can be selected from Non-Uniformity Normalized LLL (GLRLM), Dependence Non-Uniformity LHL (GLDM), Kurtosis LLL, Run Length Non-Uniformity HHL (GLRLM), Entropy LLL, Robust Mean Absolute Deviation, Sum Entropy LLL (GLCM), 90th Percentile LLL, Run Entropy HHL (GLRLM), Energy, Energy LLL, Strength (NGTDM), Autocorrelation (GLCM), Mean Absolute Deviation LLL, High Gray Level Emphasis (GLDM), Joint Average (GLCM), Sum Average (GLCM), Short Run High Gray Level Emphasis (GLRLM), Energy HHH, High Gray Level Run Emphasis (GLRLM), Run Entropy HHH (GLRLM), Energy HHL, and Mean Absolute Deviation.

[0214] The method according to any one of claims 2 to 7, wherein at least two radiomics features are selected from the radiomics features of clusters 1 to 8.

[0215] The method according to any one of claims 2 to 7, wherein at least two radiomics features are selected from the radiomics features of clusters 1 to 7.

[0216] The method according to any one of claims 2 to 7, wherein at least two radiomics features are selected from the radiomics features of clusters 1 to 6.

[0217] The method according to any one of claims 2 to 7, wherein at least two radiomics features are selected from the radiomics features of clusters 1 to 5.

[0218] Clause 13. The method of any one of Clauses 2 to 7, wherein at least two radiomics features are selected from the radiomics features of Clusters 1 to 4.

[0219] Clause 14. The method of any one of Clauses 2 to 7, wherein at least two radiomics features are selected from the radiomics features of Clusters 1 to 3.

[0220] Clause 15. The method of any one of Clauses 2 to 7, wherein the at least two radiomics features are selected from the radiomics features of Clusters 1 and 2.

[0221] Clause 16. The method of any one of Clauses 1 to 14, wherein at least two radiomics features include at least three radiomics features.

[0222] Clause 17. The method of any one of Clauses 1 to 13, wherein at least two radiomics features include at least four radiomics features.

[0223] Clause 18. The method of any one of Clauses 1 to 12, wherein at least two radiomics features include at least five radiomics features.

[0224] Clause 19. The method of any one of Clauses 1 to 11, wherein at least two radiomics features include at least six radiomics features.

[0225] Clause 20. The method of any one of Clauses 1 to 10, wherein at least two radiomics features include at least seven radiomics features.

[0226] Clause 21. The method of any one of Clauses 1 to 9, wherein at least two radiomics features include at least eight radiomics features.

[0227] Clause 22. The method of any one of Clauses 1 to 8, wherein at least two radiomics features include at least nine radiomics features.

[0228] Clause 23. The method of Clause 1, wherein at least two radiomics features comprise six radiomics features, and the six radiomics features are Short Run High Gray Level Emphasis, Skewness, Run Entropy, Small Area Low Gray Level Emphasis, Zone Entropy HHH, and Zone Entropy.

Claims

**Claim 1** A method for characterizing a perivascular region including perivascular tissue using medical imaging data of a subject, the method including calculating a value of a radiomics signature of the perivascular region using the medical imaging data, wherein the radiomics signature is calculated based on measured values of at least two radiomics features of the perivascular region, and the measured values of the at least two radiomics features are calculated from the medical imaging data, wherein the radiomics signature provides a measure of the texture of the perivascular region, Method. **Claim 2** The method according to claim 1, wherein the radiomics signature is a precursor of cardiovascular risk, and optionally, the radiomics signature is a precursor of the likelihood that the subject will experience a major adverse cardiovascular event. **Claim 3** The method according to claim 1 or 2, wherein at least one of the at least two radiomics features is calculated from a wavelet transform of attenuation values. **Claim 4** The at least two radiomics features are selected from the radiomics features of clusters 1 to 9, and each of the at least two radiomics features is selected from a different cluster, Cluster 1 consists of short run high gray level emphasis, high gray level emphasis, high gray level run emphasis, autocorrelation, sum average, joint average, and high gray level zone emphasis, 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, dependence entropy LLL, dependence entropy, zone entropy LLL, run entropy LLL, and average value LLL, Cluster 4 consists of small area low gray level emphasis, low gray level zone emphasis, short run low gray level emphasis, low gray level run emphasis, low gray level emphasis, small dependence low gray level emphasis, gray level dispersion LLL (GLSZM), gray level dispersion (GLDM), dispersion, gray level dispersion (GLDM), differential dispersion LLL, gray level dispersion LLL (GLRLM), dispersion LLL, gray level dispersion LLL (GLDM), sum of squares, contrast LLL, mean absolute deviation, interquartile range, robust mean absolute deviation, long run low gray level emphasis, differential dispersion, gray level dispersion (GLSZM), inverse difference moment normalization, mean absolute deviation LLL, sum of squares LLL, and contrast. Cluster 5 consists of zone entropy, gray level non-uniformity normalization (GLRLM), gray level non-uniformity normalization LLL (GLRLM), sum entropy, joint energy, entropy, gray level non-uniformity normalization (GLDM), joint energy, gray level non-uniformity normalization LLL (GLDM), uniformity LLL, sum entropy LLL, and uniformity. Cluster 6 consists of zone entropy HHH, size zone non-uniformity normalization HHH, and small area emphasis 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 level dispersion LLL (GLDM), dispersion LLL, gray level dispersion LLL (GLRLM), gray level dispersion (GLRLM), robust mean absolute deviation LLL, gray level dispersion (GLDM), dispersion, mean absolute deviation, cluster prominence, sum entropy LLL, interquartile range LLL, gray level dispersion LLL (GLSZM), sum of squares, robust mean absolute deviation, sum entropy, interquartile range, cluster prominence LLL, entropy LLL, 10th percentile LLL, 10th percentile. The method according to any one of claims 1 to 3, wherein the cluster 9 consists of size zone non-uniformity LLL, dependency non-uniformity HLL, gray level non-uniformity HLL (GLSZM), gray level non-uniformity (GLSZM), run length non-uniformity HHL, run length non-uniformity LHL, dependency non-uniformity LHL, dependency non-uniformity, run length non-uniformity HLH, business, run length non-uniformity LLH, dependency non-uniformity LLH, dependency non-uniformity LLL, size zone non-uniformity, energy HLL, run length non-uniformity LHH, size zone non-uniformity HLL, gray level non-uniformity LLH (GLSZM), gray level non-uniformity LHL (GLSZM), gray level non-uniformity LLL (GLSZM), run length non-uniformity HLL, gray level non-uniformity HLH (GLSZM), gray level non-uniformity HHL (GLSZM), run length non-uniformity, and run length non-uniformity HHH.

5. The method according to any one of claims 1 to 3, wherein the at least two radiomics features are selected from median LLL, mean LLL, median, root mean square LLL, mean, kurtosis, root mean square, run entropy LLL (GLRLM), uniformity, 90th percentile, gray level non-uniformity normalization (GLRLM), uniformity LLL, skewness, gray level non-uniformity normalization LLL (GLRLM), 10th percentile LLL, skewness LLL, 10th percentile, entropy, interquartile range LLL, robust mean absolute deviation LLL, run entropy (GLRLM), interquartile range, sum entropy (GLCM), gray level non-uniformity normalization LLL (GLRLM), dependency non-uniformity LHL (GLDM), kurtosis LLL, run length non-uniformity 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 level emphasis (GLDM), joint average (GLCM), sum average (GLCM), short run high gray level emphasis (GLRLM), energy HHH, high gray level run emphasis (GLRLM), run entropy HHH (GLRLM), energy HHL, and mean absolute deviation.

6. The method according to claim 4, wherein the at least two radiomics feature quantities are selected from the radiomics feature quantities of clusters 1 to 6.

7. The method according to any one of claims 1 to 6, wherein the at least two radiomics feature quantities include at least six radiomics feature quantities.

8. The method according to any one of claims 1 to 7, further comprising predicting a risk that the subject experiences a major adverse cardiac event based on at least the calculated value of the radiomics signature.

9. The method according to any one of claims 1 to 8, further comprising determining whether the subject has a vascular disease or coronary heart disease based on at least the calculated value of the radiomics signature.

10. A method for deriving a radiomics signature for predicting a cardiovascular risk, the method comprising constructing a perivascular radiomics signature for predicting a cardiovascular risk using a radiomics dataset, the radiomics signature being calculated based on at least two radiomics feature quantities of a perivascular region including perivascular tissue, the dataset includes values of a plurality of radiomics feature quantities of perivascular regions obtained from medical imaging data of each of a plurality of individuals, the plurality of individuals including a first group of individuals who reached a clinical endpoint indicating a cardiovascular risk within a subsequent period after the collection of the medical imaging data, and a second group of individuals who did not reach a clinical endpoint indicating a cardiovascular risk within the subsequent period, the radiomics signature is constructed to provide a measure of the texture of the perivascular region. Method.

11. The method according to claim 10, comprising identifying a plurality of clusters of radiomics feature quantities, each cluster including a subset of the plurality of radiomics feature quantities, each cluster including an original radiomics feature quantity selected such that each of the other radiomics feature quantities within that cluster is collinear, the at least two radiomics feature quantities each being selected from different clusters, and optionally each of the original radiomics feature quantities being selected such that it is not collinear with any of the original radiomics feature quantities of any of the other clusters.

12. The method according to claim 10 or 11, wherein the at least two radiomics features are selected such that they are not collinear with each other.

13. The method according to any one of claims 10 to 12, wherein the radiomics signature is constructed to correlate with the clinical endpoint, and optionally, the radiomics signature is constructed to be significantly associated with the clinical endpoint.

14. The method according to any one of claims 10 to 13, wherein constructing the radiomics signature is performed using a machine learning algorithm.

15. The method according to any one of claims 10 to 14, further comprising configuring a system for calculating the value of the radiomics signature for a patient.

16. The method according to any one of claims 10 to 15, further comprising characterizing the perivascular region of a patient by calculating the value of the radiomics signature for the perivascular region of the patient.

17. The method according to any one of claims 1 to 16, wherein the perivascular region comprises perivascular adipose tissue.

18. The method according to any one of claims 1 to 17, wherein the radiomics signature is calculated based on additional radiomics features of the perivascular region in addition to the at least two radiomics features.

19. A system configured to perform the method according to any one of claims 1 to 18.

Images