Radiomic signature of fat

Abstract

A method for characterizing a region of interest including adipose tissue using medical imaging data, such as a computed tomography (CT) scan based on a subject, is disclosed. The method includes using the medical imaging data to calculate a value of a radiomics signature for the region of interest. A method for deriving a radiomics signature indicative of adipose tissue dysfunction is also disclosed. The method includes obtaining a radiomics dataset and using the radiomics dataset to construct a radiomics signature for a region of interest including adipose tissue. A system for performing the foregoing method is also disclosed.

Description

Technical Field

[0001] The present invention relates to methods of characterizing adipose tissue, in particular using radiomic signatures and systems for use in said methods. The present invention also relates to methods of deriving such signatures and systems for use in said methods. Background Art

[0002] Despite the centrality of adipose tissue dysfunction in the pathogenesis of diabetes, obesity, and associated vascular disease risk, methods to study its biology are currently lacking, other than performing invasive biopsies.

[0003] Techniques for imaging adipose tissue are known (e.g., see Wang, H., Chen, Y.E. & Eitzman, DT "Imaging body fat: techniques and cardiovascular metabolic implications," Arteriosclerosis, thrombosis, and vascular biology 34, 2217-2223, doi: 10.1161 / ATVBAHA.114.303036 (2014); Christen, T. et al. "Increased glucose uptake in visceral versus subcutaneous adipose tissue revealed by PET "PET imaging of human brown adipose tissue with the TSPO tracer" [(11)C]PBR28. Molecular imaging and biology: MIB: the official publication of the Academy of Molecular Imaging, doi: 10.1007 / s11307-017-1129-z (2017). Specifically, computerized tomography is now used to image human adipose tissue, for example to quantify fat mass.

[0004] However, although human obesity can be reliably described and quantified using volumetric methods, it is indirectly or poorly linked to adipose tissue biological features. Adipose tissue dysfunction plays a central role in the development of insulin resistance and obesity-related vascular disease. However, expansion of fat per se does not always lead to the development of insulin resistance, which depends on various aspects of adipose tissue biology, such as adipocyte hyperplasia or hypertrophy, adipogenic capacity, adipose tissue inflammation, and fibrosis. As a result, adipose tissue biological phenotypes cannot be accurately captured by simple volumetric quantification of human obesity.

[0005] The average radiodensity of visceral or subcutaneous fat has recently been proposed as a simple measure of adipose tissue mass, which is independently associated with cardiovascular risk in clinical cohorts (see Rosenquist, KJ et al. "Visceral and subcutaneous fat quality and cardiometabolic risk" JACC. Cardiovascular Imaging 6, 762-771, doi: 10.1016 / j.jcmg.2012.11.021 (2013)). Other methods (such as PET / CT imaging) can be used to study adipose tissue metabolic activity (e.g., by quantitative 18F-FDG uptake), but PET is limited by its availability, cost, and high radiation exposure.

[0006] To date, there are no methods to adequately phenotype adipose tissue biology by noninvasive means. Therefore, there is an unmet need for noninvasive methods to detect adipose tissue dysfunction and assess obesity-related metabolic burden. Summary of the invention

[0007] According to a first aspect of the invention, a method for characterizing a region of interest comprising adipose tissue (e.g., its phenotype, e.g., composition and / or texture) is provided. The method may be used to identify or diagnose adipose tissue dysfunction, or to identify phenotypic characteristics of adipose tissue dysfunction. The method may include using medical imaging data to calculate a value of a radiomic signature of the region of interest. The radiomic signature may be calculated based on measurements of a plurality of radiomic features of the region of interest. The measurements of the radiomic features may be calculated based on the medical imaging data.

[0008] The radiomics signature may provide a measure of texture of the region of interest or adipose tissue. At least one of the radiomics features may provide a measure of texture of the region of interest or adipose tissue. For example, at least one of the radiomics features may be a texture statistic.

[0009] The radiomic signature (i.e., its value) may indicate or be associated with (e.g., statistically significantly associated with) adipose tissue dysfunction, specifically adipose tissue in a region of interest. Alternatively, the radiomic signature may indicate or be associated with a surrogate marker of adipose tissue dysfunction, such as a metabolic disorder associated with adipose tissue dysfunction, a phenotypic characteristic of adipose tissue dysfunction, or a gene expression profile associated with adipose tissue dysfunction.

[0010] The radiomic signature (ie, its value) can predict the likelihood that a subject will suffer from a metabolic disorder, in particular a metabolic disorder associated with adipose tissue dysfunction (eg, diabetes or insulin resistance).

[0011] The plurality of radiomic features may include at least two radiomic features selected from the radiomic features of clusters 1 to 5, wherein the at least two radiomic features are each selected from a different group, and wherein:

[0012] Group 1 consisted of: large dependence low gray level emphasis, sum mean, joint mean, high gray level emphasis, short run high gray level emphasis, autocorrelation, and high gray level run emphasis;

[0013] Group 2 consists of: total energy HHL, energy HHL, size zone non-uniformity HHH, size zone non-uniformity HLH, size zone non-uniformity LHH, normalized gray level non-uniformity HHL (GLDM), inverse variance HHL, normalized gray level non-uniformity HHL (GLSZM), size zone non-uniformity LHL, run length non-uniformity HHH, size zone non-uniformity HLL, dependence non-uniformity HHL, normalized gray level non-uniformity HHH (GLDM), inverse moment HHH, inverse difference HHH, non-uniformity HHH, related information measure 2, inverse variance HHH, total energy LHL, size zone non-uniformity, sphericity, size zone non-uniformity LLH, normalized run length non-uniformity H HL, normalized gray level non-uniformity HHH (GLSZM), normalized dependency non-uniformity HHL, zone percentage HHL, small dependency focus HHL, normalized size zone non-uniformity HHL, small area focus HHL, correlation, dependency non-uniformity LHH, total energy LHH, short run focus HHH, normalized run length non-uniformity HHH, dependency non-uniformity LHL, inverse difference LHH, inverse moment LHH, small dependency focus HHH, dependency variance HLH, run percentage HHH, zone percentage HHH, inverse difference LHL, dependency non-uniformity HLH, inverse variance LHH, large dependency focus HLH, normalized run length non-uniformity HLH, short run focus HLH, Normalized dependency non-uniformity HLH, difference variance, long run focus HLH, energy LHL, contrast (GLCM), run percentage HLH, joint entropy LHH, difference entropy, small dependency focus HLH, normalized size zone non-uniformity HLH, small area focus HLH, difference entropy LHH, difference mean, run variance HLH, inverse difference moment, inverse difference, inverse variance, run entropy LHH, entropy LHH, zone percentage HLH, sum entropy LHH, normalized inverse difference moment, normalized inverse difference, small dependency low gray level focus, energy LHH, joint entropy LHL, small area low gray level focus, intensity, dependency non-uniformity LLH, run length non-uniformity HHL, normalized gray level non-uniformity LHL (GLDM), Busyness, Normalized Dependency Non-uniformity LHH, Small Dependency Focus LHH, Coarseness, Zone Percentage LHH, Normalized Dependency Non-uniformity LHL, Run Length Non-uniformity HLH, Small Dependency Focus LHL, Joint Entropy, Small Area Focus LHL, Normalized Size Zone Non-uniformity LHL, Short Run Focus HLL, Run Length Non-uniformity LHH, Normalized Run Length Non-uniformity HLL, Large Dependency Focus HLL, Long Run Focus HLL, Run Percentage HLL, Dependency Variance HLL, Dependency Non-uniformity HLL, Zone Percentage HLL, Dependency Non-uniformity, Dependency Non-uniformity HHH, Normalized Dependency Non-uniformity LLH, Low Gray Level Zone Focus, and Short Run Low Gray Level Focus;

[0014] Group 3 consists of: dependency entropy, zone entropy, correlation, correlation information measure 2, zone percentage HHH, small dependency focus HHH, run percentage HHH, normalized dependency non-uniformity LHL, small dependency focus LHL, zone percentage HHL, small dependency high gray level focus, small dependency focus HHL, normalized dependency non-uniformity HHL, short run focus HHH, normalized run length non-uniformity HHH, normalized run length non-uniformity HHL, normalized size zone non-uniformity LHL, small area focus LHL, Normalized size zone non-uniformity HHL, small area focus HHL, zone percentage HLH, run variance HLH, small dependence focus HLH, long run focus HLH, normalized size zone non-uniformity LLH, small area focus LLH, normalized size zone non-uniformity HLH, small area focus HLH, run percentage HLH, normalized run length non-uniformity HLH, short run focus HLH, large dependence focus HLH, zone percentage LHH, small dependence focus LLH, small dependence focus LHH, normalized dependence non-uniformity HLH, zone percentage LLH, dependent variance HLH, normalized dependent non-uniformity LLH, normalized dependent non-uniformity LHH, inverse difference LHL, inverse difference moment HHH, inverse difference HHH, uniformity HHH, normalized gray level non-uniformity LHL (GLSZM), normalized gray level non-uniformity HHH (GLDM), normalized gray level non-uniformity HHL (GLDM), joint entropy LHL, normalized inverse difference, normalized inverse difference moment, long run focus HLL, zone percentage HLL, inverse difference, inverse difference moment, Difference mean, run percentage HLL, large dependence focus HLL, contrast (GLCM), short run focus HLL, normalized run length non-uniformity HLL, inverse variance, inverse variance HHL, difference entropy, dependence variance HLL, inverse difference moment LHH, difference variance, entropy LHH, inverse variance LHH, inverse difference LHH, inverse variance HHH, sum entropy LHH, joint entropy LHH, normalized gray level non-uniformity HHH (GLSZM), normalized gray level non-uniformity HHL (GLSZM), difference entropy LHH and run entropy LHH;

[0015] Group 4 consists of mean LLL, median LLL, and energy LLL; and

[0016] Group 5 consists of run-length entropy.

[0017] The at least two radiomic features may include at least two of the following: large dependency low gray level emphasis, total energy HHL, dependency entropy, mean LLL, and run entropy. The at least two radiomic features may be composed of five radiomic features, and the five radiomic features may be composed of the following: large dependency low gray level emphasis, total energy HHL, dependency entropy, mean LLL, and run entropy.

[0018] The at least two radiomic features may be selected from the radiomic features of clusters A to C, wherein the at least two radiomic features are each selected from a different cluster, and wherein:

[0019] Cluster A consisted of the radiomic features of groups 1 to 3;

[0020] Cluster B consists of the radiomics features of Group 4; and

[0021] Cluster C consisted of the radiomic features of group 5.

[0022] The plurality of radiomic features may include at least two radiomic features selected from the radiomic features of clusters A to C, wherein the at least two radiomic features are each selected from a different cluster, and wherein:

[0023] Cluster A is composed of: large dependence low gray level emphasis, run entropy, dependence entropy, cluster shading, skewness, run variance HLH, number of voxels, and gray level non-uniformity;

[0024] Cluster B consists of: mean LLL, median LLL, joint mean, median, complexity, long run high gray level focus, sphericity, kurtosis, and roughness HHH; and

[0025] Cluster C consists of the following: major axis, small dependent high gray level focus, minor axis, energy LLL, maximum 2D diameter row, long run low gray level focus, total energy HHL, normalized dependent non-uniformity HHH, contrast (GLCM), surface to volume ratio, and entropy, normalized size area non-uniformity HHH, cluster prominence LHH, contrast LLH (GLCM), and energy LHL.

[0026] The multiple radiomic features may include at least two radiomic features selected from the following: large dependence low gray level emphasis, run entropy, dependence entropy, cluster shading, skewness, run variance HLH, number of voxels, gray level non-uniformity, mean LLL, median LLL, joint mean, median, complexity, long run high gray level emphasis, sphericity, kurtosis, coarseness HHH, major axis, small dependence high gray level emphasis, minor axis, energy LLL, maximum 2D diameter row, long run low gray level emphasis, total energy HHL, normalized dependence non-uniformity HHH, contrast (GLCM), surface volume ratio, and entropy, normalized size area non-uniformity HHH, cluster prominence LHH, contrast LLH (GLCM) and energy LHL.

[0027] The at least two radiomic features may include at least three radiomic features. The at least two radiomic features may include at least four radiomic features. The at least two radiomic features may include at least five radiomic features.

[0028] The medical imaging data may include an attenuation value for each voxel of a plurality of voxels corresponding to at least a region of interest.

[0029] At least one of the plurality of radiomic features may be calculated based on a wavelet transform of the attenuation values.

[0030] The method may further include identifying the region of interest based on or using the medical imaging data. The region of interest may be identified using a manual outline. The region of interest may be identified as a voxel containing only the medical imaging data, the voxel having an attenuation value that falls within a given attenuation value range. The given range may be from approximately -190 Hounsfield Units to approximately -30 Hounsfield Units.

[0031] The method may further include segmenting the region of interest. The method may further include calculating a value of the radiomics feature based on the segmented region of interest.

[0032] The method may further comprise predicting the subject's risk of suffering from a metabolic disorder based at least on the calculated value of at least the radiomic signature. The metabolic disorder may be diabetes or insulin resistance.

[0033] The method may further include determining or diagnosing whether the subject suffers from adipose tissue dysfunction or a phenotypic characteristic of adipose tissue dysfunction (eg, fibrosis or inflammation) based on at least the calculated value of the radiomic signature.

[0034] The radiomic signature may include a weighted sum of the plurality of radiomic features. The radiomic signature may be linearly related to the weighted sum of the radiomic features.

[0035] According to a second aspect of the present invention, a method for deriving a radiomic signature is provided. The radiomic signature may indicate adipose tissue dysfunction. The radiomic signature may be suitable for identifying or diagnosing adipose tissue dysfunction. The method may include using a radiomic dataset to construct a radiomic signature indicating adipose tissue dysfunction. The radiomic signature may be calculated based on a second plurality of radiomic features. The dataset may include values ​​of a first plurality of radiomic features obtained from medical imaging data of a region of interest, the region of interest including adipose tissue of each individual in a plurality of individuals. The plurality of individuals may include a first group of individuals identified as having adipose tissue dysfunction (particularly when collecting or recording medical imaging data) and a second group of individuals identified as not having adipose tissue dysfunction (particularly when collecting or recording medical imaging data). The second plurality of radiomic features may be selected from the first plurality of radiomic features, specifically as determined according to the dataset or using the dataset (e.g., using a machine learning algorithm), to provide a radiomic signature indicating adipose tissue dysfunction.

[0036] Based at least in part on the markers of adipose tissue dysfunction, each of the individuals may be identified as having or not having adipose tissue dysfunction.The radiomics dataset may further include markers of adipose tissue dysfunction.

[0037] For each of the plurality of individuals, the marker may be or may include or may indicate the presence or absence of a metabolic disorder associated with adipose tissue dysfunction, such as diabetes or insulin resistance (i.e., information that distinguishes whether each of the individuals has or does not have a metabolic disorder).

[0038] Markers may include phenotypic characteristics of adipose tissue associated with adipose tissue dysfunction, such as fibrosis or inflammation.

[0039] The marker may include a surrogate marker of a phenotypic property of adipose tissue associated with adipose tissue dysfunction, such as a selected gene expression profile. Specifically, the surrogate marker may be or may include a gene expression profile that is a marker of a phenotypic property of adipose tissue dysfunction.

[0040] The method may further include identifying a first subset of the first plurality of radiomic features that are not collinear with each other as determined according to the data set. The second plurality of radiomic features may include at least two radiomic features, each of the at least two radiomic features being selected to belong to, be collinear with, or be highly correlated with a different radiomic feature of the first subset (as determined according to the data set).

[0041] The method may further include identifying a second subset of radiomic features from the first subset on which the radiomic signature should be calculated (i.e., the best features) using a first feature selection algorithm (e.g., a first machine learning feature selection algorithm). In other words, for example, when the radiomic signature is calculated based on the second subset of radiomic features (which has been optimized), the second subset of radiomic features is predicted to maximize the accuracy of the radiomic signature (e.g., optimize or maximize the association of the radiomic signature with adipose tissue dysfunction). The second subset of radiomic features can maximize the association of the preliminary radiomic signature with adipose tissue dysfunction (when the preliminary radiomic signature is calculated based on the second subset of radiomic features), as determined according to the data set. The at least two radiomic features may be selected as different radiomic features belonging to the second subset or collinear therewith (as determined according to the data set). The at least two radiomic features may include all radiomic features belonging to the second subset, or their collinear equivalents, which are collinear with the radiomic features, as determined according to the data set. In other words, the at least two radiomic features may include each of the radiomic features belonging to the second subset or radiomic features collinear therewith.

[0042] Each radiomic feature of the at least two (or second plurality) of radiomic features may be selected as stable as determined from the data set. Prior to selecting the second plurality of radiomic features, all unstable radiomic features may be removed from the first plurality of radiomic features.

[0043] If the intraclass correlation coefficient (e.g., scan-re-scan and / or multiple delineation) for the radiomics feature is less than a stability threshold (e.g., as determined from the data set), the radiomics feature can be identified as unstable. The stability threshold can be at least about 0.9.

[0044] The method may further include identifying a plurality of clusters of the first plurality of radiomic features (e.g., a first subset of radiomic features) by performing a cluster analysis, such as a correlation cluster analysis (e.g., using a clustering algorithm, specifically a machine learning clustering algorithm). The second plurality or at least two radiomic features may each be selected from a different cluster or selected to be collinear with radiomic features from different clusters. The cluster analysis may identify clusters based on the strength of correlations between radiomic features. Intra-cluster correlations may be stronger than inter-cluster correlations. For example, the degree of correlation of radiomic features within each cluster with each other may be greater than their degree of correlation with radiomic features in other clusters.

[0045] The cluster analysis can be a hierarchical cluster analysis, a k-means cluster analysis, a distribution-based cluster analysis, or a density-based cluster analysis. Specifically, the cluster analysis can be a hierarchical cluster analysis. For example, in a correlation graph, a cluster analysis, such as a clustering algorithm, can identify clusters based on distances between radiomic features, such as squared Euclidean distances between radiomic features. A clustering algorithm can identify clusters based on distances between features in a correlation space, where the distance between each pair of features corresponds to the degree to which those features are correlated, i.e., the closer two features are in the correlation space, the more correlated they are to each other.

[0046] The method may include identifying multiple clusters of radiomic features. Each cluster may include a subset of the multiple radiomic features. Each cluster may include an original radiomic feature, using which each of the other radiomic features in the cluster is selected to be collinear, for example, as determined or calculated based on the data set. At least two radiomic features may each be selected from different clusters.

[0047] Each of the raw radiomic features may be selected to be non-collinear with any of the raw radiomic features of any of the other clusters, eg, as determined or calculated from the data set.

[0048] Each of the radiomic features in each cluster can be selected to be collinear with all other radiomic features in the same cluster, for example, as determined or calculated from the data set.

[0049] Each of the raw radiomic features can be selected to be significantly associated with a clinical endpoint, eg, as determined or calculated from the data set.

[0050] Each of the raw radiomic signatures can be selected to be closely associated with the clinical endpoint of all radiomic signatures in its cluster, e.g., as determined or calculated from the dataset.

[0051] At least two radiomic features can be selected to be non-collinear with each other, for example, as determined or calculated from the data set.

[0052] The method may include identifying a plurality of groups of radiomic features. Each group may include a subset of the first plurality of radiomic features. Each group may include raw radiomic features (e.g., one of the first subset) using which each of the other radiomic features in the group is selected to be collinear, e.g., as determined or calculated from a data set. A second plurality of radiomic features, e.g., at least two radiomic features, may each be selected from a different group.

[0053] Each of the raw radiomic features may be selected to be non-collinear with any of the raw radiomic features of any of the other groups, eg, as determined or calculated from the data set.

[0054] Each of the radiomic features in each group can be selected to be collinear with all other radiomic features in the same group, for example, as determined or calculated from the data set.

[0055] The second plurality or at least two radiomic features may be selected to be non-collinear with each other, for example, as determined or calculated from the data set.

[0056] If the degree of correlation of two radiomic features is at least equal to a correlation threshold, they can be identified as collinear with each other, as determined from the data set. The correlation threshold can be calculated using Spearman's rho coefficient. The correlation threshold can be at least about |rho|=0.75, for example, at least about |rho|=0.9. If the degree of correlation of two radiomic features is at least about |rho|=0.75, they can be identified as highly correlated.

[0057] The radiomic signature can be constructed to be associated with adipose tissue dysfunction, as determined from a data set. For example, the radiomic signature can be constructed to be significantly associated with adipose tissue dysfunction. If a radiomic signature is associated with adipose tissue dysfunction above a significance threshold, it can be identified as being significantly associated with adipose tissue dysfunction. The significance threshold is at least about α=0.05. The association of the radiomic signature with AT dysfunction can be calculated based on or using a receiver operating characteristic (ROC) curve analysis, specifically using an area under the curve (AUC) measurement (i.e., a C statistic).

[0058] The step of constructing the radiomic signature may include refining the contribution of each radiomic feature in the second plurality of radiomic features to the radiomic signature to increase the association of the radiomic signature with adipose tissue dysfunction as determined from the dataset.

[0059] The dataset may be divided into a training cohort dataset and a validation cohort dataset, and the step of constructing the radiomics signature may include deriving the signature using at least the data for the training cohort, and validating the signature using the data for the validation cohort.

[0060] The step of constructing the radiomic signature may be performed using a machine learning algorithm. The method may include selecting a second plurality of radiomic features from the first plurality of radiomic features, specifically from the first subset, using a machine learning algorithm to provide a radiomic signature indicative of adipose tissue dysfunction, as determined based on the data set.

[0061] The step of constructing the radiomics signature may be performed using multiple cross validation. The step of constructing the radiomics signature may be performed using recursive feature elimination, for example using a random forest algorithm. The step of constructing the radiomics signature may be performed using logistic regression.

[0062] For example, the step of constructing a radiomics signature may include, for example, using a feature selection machine learning algorithm to identify a subset of radiomic features to be included in the signature. Feature selection may be performed using recursive feature elimination, for example, using a random forest algorithm. Then, the step of constructing a radiomics signature may include refining or optimizing the radiomics signature based on a subset (e.g., a second subset) of the radiomics features identified in the feature selection step, for example, using logistic regression, again optionally using a machine learning algorithm. For example, a second machine learning algorithm (e.g., a radiomics signature optimization algorithm, such as a logistic regression algorithm) may be used to refine or optimize the radiomics signature calculated using a subset of radiomics features, for example, using multiple cross validation. In other words, the second machine learning algorithm refines or optimizes the radiomics signature calculated based on a subset of radiomics features.

[0063] The radiomic signature may include a weighted sum of the second plurality of radiomic features. The radiomic signature may be linearly related to the weighted sum of the second plurality of radiomic features.

[0064] The step of constructing the radiomic signature may include adjusting the relative weight of each radiomic feature in the second plurality of radiomic features to increase the association of the radiomic signature with adipose tissue dysfunction as determined based on the data set.

[0065] Radiomic signatures can be constructed to provide a measure of the texture of a region of interest.

[0066] At least one of the second plurality of radiomic features (e.g., one of the at least two radiomic features) may provide a measure of texture of the region of interest. For example, at least one of the second plurality of radiomic features may be a texture statistic. For example, each of the second plurality of radiomic features may provide a measure of texture of the region of interest (i.e., each of the radiomic features may be a texture statistic).

[0067] The method may further include configuring a system for calculating the value of the radiomic signature of the patient, specifically for a region of interest of the patient including adipose tissue. For example, the method may further include configuring a system for characterizing a region of interest of a patient or subject by calculating the value of the radiomic signature derived for the patient or subject. The system may be configured to calculate the value of the derived radiomic signature using or based on medical imaging data of at least one region of interest of the patient or subject. The system may be configured to calculate the value of the derived radiomic signature using or based on at least a second plurality of radiomic features of a region of interest of the patient or subject.

[0068] Thus, the method may be a system for deriving a radiomic signature and configured to use the derived radiomic signature to characterize a region of interest of a patient, including adipose tissue.

[0069] The system may be configured to receive as input the medical imaging data or the value of the second plurality of radiomic features. The system may be configured to output (e.g., display) the calculated value of the radiomics label or a value based on the calculated value of the radiomics label. The system may be configured to output an indication of whether the patient suffers from adipose tissue dysfunction. The system may be configured to output an indication of the risk of the patient suffering from a metabolic disorder. The system may be a computer system.

[0070] The method may include providing instructions for configuring a system for calculating a value of the derived radiomics signature for a patient or subject.

[0071] The method may further include calculating the value of the derived radiomics signature for the area of ​​interest of the patient or subject. For example, the method may further include characterizing the area of ​​interest of the patient or subject by calculating the value for the derived radiomics signature. The value of the derived radiomics signature may be calculated based on or using medical imaging data of at least the area of ​​interest of the patient or subject. The value of the derived radiomics signature may be calculated using or at least based on the value of a second plurality of radiomics features of the area of ​​interest of the patient or subject. The area of ​​interest of the patient includes adipose tissue.

[0072] Thus, the method may be for deriving a radiomic signature and for characterizing a region of interest using the derived radiomic signature.

[0073] The medical imaging data may be radiographic data. The medical imaging data may be computed tomography data.

[0074] The adipose tissue may be subcutaneous adipose tissue. The adipose tissue may be visceral adipose tissue. The adipose tissue may be chest adipose tissue.

[0075] Adipose tissue can include or consist of non-cardiovascular adipose tissue (i.e., non-cardiac or non-vascular adipose tissue). For example, adipose tissue can include or consist of non-epicardial, non-pericardial, or non-perivascular adipose tissue. For example, adipose tissue can include a majority of non-cardiovascular adipose tissue (i.e., greater than about 50% by volume). For example, adipose tissue can include substantially non-cardiovascular adipose tissue. For example, adipose tissue can include less than about 10% (by volume) of cardiovascular adipose tissue, specifically less than about 5% of cardiovascular adipose tissue.

[0076] Alternatively, the region of interest may be distal to (i.e., not adjacent to or attached to) the cardiovascular system. For example, the region of interest may be distal to cardiovascular organs such as the heart and blood vessels, specifically arteries and veins. The adipose tissue may not be cardiac or vascular adipose tissue. For example, the adipose tissue may not be epicardial, pericardial, or perivascular adipose tissue.

[0077] In addition to the at least two radiomic features mentioned above, the radiomic signature of the present invention can also be calculated based on other radiomic features of the region of interest. For example, the radiomic signature may include other radiomic features in addition to the at least two radiomic features. Therefore, it can be said that the radiomic signature is calculated based on multiple radiomic features, and the multiple radiomic features may include at least two radiomic features.

[0078] The method of the present invention may further comprise the step of calculating a radiomics feature based on the medical imaging data.

[0079] According to a third aspect of the present invention, there is provided a system configured to perform any of the methods described above. The system may be a computer system. The system may include a processor configured to perform the steps of the method. The system may include a memory loaded with executable instructions for performing the steps of the method.

[0080] According to a fourth aspect of the invention, there is provided a use of a radiomic signature for any of the purposes described above, such as for characterizing a region of interest, detecting adipose tissue dysfunction or predicting the risk of developing a metabolic disorder. The radiomic signature can be calculated based on the measurements of a plurality of radiomic features of the region of interest.

[0081] The medical imaging data may be radiographic data. The medical imaging data may be computed tomography data. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The present invention will now be described with reference to the accompanying drawings, in which:

[0083] Figure 1 Methods for deriving radiomic signatures are outlined. A total of 225 patients undergoing coronary artery bypass grafting were included in the analysis phase. Non-contrast enhanced computed tomography (CT) images were used to link adipose tissue radiomic features to tissue biology. Only stable radiomic features (i.e., those that showed an intraclass correlation coefficient (ICC) ≥ 0.90 during the training phase) were included in the analysis. Radiomic features associated with adipose tissue biology were internally validated using a machine learning algorithm with five-fold cross-validation. The main radiomic features were then externally validated in Arm 2 as a target for 18 Biomarkers of adipose tissue inflammation imaged by F-FDG positron emission tomography / computed tomography (PET / CT) and external validation of the main radiomics signature as a biomarker of vascular disease progression in panel 3. AT: adipose tissue, ScAT: subcutaneous adipose tissue.

[0084] Figure 2 shows aspects of the principal component analysis of the adipose tissue radiomics signature. Figure 2(a) is a steep slope plot of the percentage of variation explained by the first 86 principal components, which account for 99.5% of the variation in panel 1 (843 radiomics signatures in 225 patients). Figure 2(b) is a component plot of the three main principal components. Figure 2(c) is a correlation plot of the first three components with demographic, biochemical characteristics (n=225), and gene expression profile patterns (n=167) in adipose tissue of panel 1 patients.

[0085] Figure 3 illustrates aspects of the unsupervised hierarchical clustering of patients and radiomic features. Figure 3(a) illustrates the unsupervised hierarchical clustering of group 1 patients using 196 stable radiomic features (after stability assessment) in the group 1 population. The X-axis represents the different radiomic features and the Y-axis represents the number of observations (n=225 patients). The row dendrogram represents two different clusters of patients. Figure 3(b) is a similarity-dissimilarity plot showing the Euclidean distance (i.e., dissimilarity) between selected radiomic features (n=196 after stability assessment) for visual identification of possible clusters in the group 1 cohort of patients. Figure 3(c) is a correlation plot of 32 filtered (non-collinear) radiomic features after excluding highly correlated (i.e., collinear) features (|rho|>0.90). P value of chi-square.

[0086] FIG4 illustrates the detection of adipose tissue dysfunction by radiomic phenotype using radiomic signatures. FIG4(a) and (b) illustrate the association of adipose tissue gene expression profiles (FABP4 as a surrogate marker of adipocyte differentiation and TNFA as a surrogate marker of inflammation) with diabetes and insulin resistance (HOMO-IR) in non-diabetic patients, respectively. FIG4(c) and (d) illustrate the sensitivity of multivariate logistic regression models to high FABP4 and TNFA expression in adipose tissue of the same patients, respectively, demonstrating the incremental discriminative value of radiomic signatures for detecting adipocyte differentiation and adipose tissue inflammatory states, respectively. FIG4(e) illustrates aspects of radiomic signature construction by machine learning using recursive feature elimination (RFE) with five-fold cross validation and selection of key radiomic signatures for detecting adipose tissue dysfunction. FIG4(f) demonstrates the incremental diagnostic value of key radiomic signatures for detecting adipose tissue dysfunction by clinical risk factors, obesity index, and homeostatic model of insulin resistance (HOMA-IR).

[0087] FIG5 shows that a radiomic signature of a region of interest (in this case, the subcutaneous region) including adipose tissue can be used as a radiomic biomarker of adipose tissue health and metabolic risk. FIG5(a) shows the association of the radiomic signature of the present invention (referred to as "FatHealth") with adipose tissue differentiation / inflammatory phenotype based on the expression levels of FABP4 and TNFA. FIG5(b) plots a logistic regression for classifying dysfunctional adipose tissue by the radiomic signature (FatHealth). The association between the number of cardiometabolic risk factors (i.e., smoking, hypertension, diabetes) and FatHealth in the derivation and test cohorts of group 1 patients is shown using the box and whisker plots in FIG5(c) and (d), respectively. FIG5(e) shows the external validation of the radiomic signature (FatHealth) of subcutaneous adipose tissue uptake of 18-fluorodeoxyglucose in PET / CT imaging, and FIG5(f) shows the diagnostic accuracy of the radiomic signature for high adipose tissue inflammation (defined as the ratio of tissue to background in the highest tertile). AUC: area under the curve; TBR: target to background ratio. DETAILED DESCRIPTION

[0088] The inventors have discovered that a radiomic signature (also referred to as a "score" or "index") calculated based on two or more radiomic features of a region of interest (ROI) including adipose tissue is independently associated with different biological phenotypes of adipose tissue (AT) that are associated with AT dysfunction. Generally, as used herein, "correlation" may be understood to mean "statistically correlated", such as "statistically significantly correlated". Specifically, the radiomic signature is able to distinguish the expression of gene markers that are characteristic of AT dysfunction, such as differentiation of adipocytes and adipose tissue inflammation. Therefore, the radiomic signature of the present invention is preferably calculated based on two or more radiomic features of a region of interest including AT, and provides a tool for characterizing a region of interest (specifically fat or adipose) and for non-invasive detection or identification of adipose tissue dysfunction.

[0089] The radiomic signatures of the invention can be used alone to characterize the region of interest to provide diagnostic or prognostic information, or they can be used in combination with existing models and risk factors, such as age, sex, hypertension, dyslipidemia, smoking, diabetes, body mass index, homeostasis model assessment of insulin resistance (HOMA-IR) index, and / or subcutaneous fat mass.

[0090] The present invention utilizes the following discovery: specific imaging patterns in adipose tissue (e.g., related to tissue texture and radiodensity) are associated with different biological phenotypes and gene markers of adipose tissue characteristics or phenotypes indicating AT dysfunction (such as adipocyte differentiation and adipose tissue inflammation). The radiotranscriptomics method of the present invention provides a means for non-invasively detecting, identifying or diagnosing adipose tissue dysfunction, which has only been feasible so far by invasive means such as fat biopsy. The radiomics signature of the present invention can be used to assess adipose tissue health or identify adipose dysfunction, capturing the impact of risk factors on adipose tissue biology, which is valuable in the fields of endocrinology and diabetes. Therefore, the radiomics signature can be used as a measure of metabolic risk (e.g., cardiometabolic risk) or predict metabolic risk. For example, the radiomics signature of the present invention can be used to classify patients according to their metabolic risk, that is, their risk of developing metabolic disorders, such as diabetes (diabetes), specifically type II diabetes. The radiomics signature of the present invention can be used to evaluate the response to therapeutic intervention. For example, the radiomics signature can be calculated before a round of treatment for a metabolic disorder, and then after the round of treatment, the signature difference is used to evaluate the response to treatment.

[0091] The radiomic signature of the present invention exploits the finding that, in addition to simple fat mass measurements, fat texture (i.e., spatial heterogeneity or variability) is associated with metabolic risk and features of adipose tissue dysfunction. Thus, the radiomic signature of the present invention can be constructed to provide a measure of the texture of a region of interest. Thus, if the ROI is a subcutaneous region, the radiomic signature of the present invention can also be referred to as a texture index, such as a subcutaneous texture index (STI).

[0092] The region of interest (ROI) is an area or volume of an individual's (i.e., patient's) body that includes adipose tissue. Thus, the ROI may be an area or volume of AT, or may include or consist of AT. The adipose tissue may be subcutaneous adipose tissue (ScAT), visceral (or abdominal) adipose tissue, or chest adipose tissue. Specifically, the adipose tissue may be subcutaneous adipose tissue, which is minimally affected by the biology of neighboring organs. Subcutaneous tissue, also known as hypodermis, is tissue located near and below the skin. Tissue is a complex biological structure and may include cells (e.g., adipocytes, neurons, etc.) as well as extracellular structures and substances (such as water) that may occupy intercellular spaces. For example, subcutaneous adipose tissue typically includes fibroblasts, adipocytes, and macrophages. Visceral adipose tissue, sometimes referred to as abdominal or intra-abdominal adipose tissue, is located in the abdominal cavity between the intra-abdominal organs. Chest adipose tissue is located in the chest (or chest) cavity.

[0093] Preferably, the ROI includes or consists of non-cardiovascular adipose tissue. For example, the adipose tissue may include a large portion of non-cardiovascular adipose tissue (i.e., based on the total volume of the adipose tissue, greater than 50% by volume). For example, the adipose tissue may include substantially non-cardiovascular adipose tissue. For example, the adipose tissue may include less than 10% (by volume, based on the total volume of the adipose tissue) of cardiovascular adipose tissue, specifically less than 5% (by volume) of cardiovascular adipose tissue. Adipose tissue may be defined as all voxels in the ROI with a radiodensity within a given range (e.g., about -190Hu to about -30HU). Cardiovascular adipose tissue is adipose tissue associated with the cardiovascular system (i.e., adjacent to, attached to, or part thereof). Specifically, cardiovascular adipose tissue includes cardiac (i.e., heart) fat and vascular fat. For example, cardiovascular fat may refer to epicardium, pericardium, and perivascular adipose tissue. Therefore, the region of interest may not contain cardiovascular adipose tissue, specifically the region of interest may not contain epicardium, pericardium, or perivascular adipose tissue, or may include adipose tissue other than epicardium, pericardium, or perivascular adipose tissue. Thus, the ROI may not be adjacent to (i.e., may be distant from or spaced apart from) an organ of the cardiovascular system (i.e., a cardiovascular organ such as the heart or major blood vessels such as arteries and veins). Cardiovascular adipose tissue is affected by the underlying health of the cardiovascular system, specifically by the phenotype of the organ tissue to which it is attached (e.g., myocardial or vascular tissue). Therefore, if the ROI includes or consists of non-cardiovascular adipose tissue, the signature of the present invention is a more reliable indicator of metabolic health and adipose tissue dysfunction associated with metabolic health, which reflects metabolic health and the health of fat itself more directly than cardiovascular fat.

[0094] The present invention employs a radiomics approach. Radiomics is an imaging field in which a large amount of quantitative information is extracted from imaging data using data characterization algorithms. The resulting features, referred to as radiomic features, range from simple volume, shape-related or first-order statistics such as mean or median attenuation to second- and higher-order statistics that describe the texture of segmented volumes or regions and the spatial relationship of voxels with similar or different attenuation values. Such features can identify clinically important imaging patterns that are not discernible to the naked eye and have the potential to maximize the diagnostic yield of non-invasive AT phenotypes.

[0095] The labels of the present invention are derived and calculated based on radiomic features, such as radiomic features extracted from medical imaging data. Specifically, the medical imaging data from which the radiomic features are extracted corresponds to a region of interest (ROI), and optionally also corresponds to other tissues adjacent to or surrounding the ROI, such as skin. Medical imaging data typically includes radiodensity (or attenuation) values ​​for a plurality of voxels of a region of interest (in this case, the ROI), typically expressed in Hounsfield units (HU), and optionally also includes adjacent tissues.

[0096] The medical imaging data is preferably computed tomography (CT) data, but other forms of medical imaging data (e.g., radiographic data) that provide attenuation (or radiodensity) data for voxels of the imaging region, such as three-dimensional computed tomography data, may be used instead. Typically, the medical imaging data used in the present invention is three-dimensional imaging data. Throughout the following, where reference is made to CT or another medical imaging technique, it should be understood that other suitable medical imaging techniques may be used alternatively.

[0097] ROI can be identified by people such as operators, for example, by manual contour (delineation). Operators can identify ROI by checking imaging data (such as CT images). Alternatively, or in addition to using manual contour to identify the spatial identification of ROI, ROI can also be identified by applying a ray density (or attenuation) mask to the data and identifying ROI as only those voxels with radiodensity included in a given or predetermined range. For example, ROI can only be included in those voxels with a ray density of about -190HU to about +30HU within the range of Hounsfield units and optionally also located in the contour area identified by the operator. Specifically, ROI can be defined as all voxels with a radiodensity of about -190Hu to about -30HU within the range of Hounsfield units. This range of attenuation values ​​generally corresponds to the radiodensity of adipose tissue. However, other ranges can be used or included, such as about -30 to about +30 Hounsfield units, which generally correspond to the radiodensity of water.

[0098] The ROI may be segmented prior to calculating the radiomic features and the radiomic features calculated from the segmented data. The segmented volume or region corresponds to the ROI, and the segmentation may remove data corresponding to voxels outside the ROI. Thus, segmentation may be achieved by identifying the ROI as described above, and then removing any voxels from the data identified as not being part of the ROI, such as those corresponding to surrounding or adjacent tissue voxels. The segmented ROI may then be extracted and used to calculate the radiomic features.

[0099] Computer programs or software can be used to perform calculations of radiomic features based on medical imaging data. For this purpose, there are various commercially available software packages, such as 3D Slicer (available from http: / / www.slicer.org; see Fedorov, A. et al. "3D Slicer as an image computing platform for the Quantitative Imaging Network" "Magnetic Resonance Imaging (Magn Reson Imaging)" 30,1323-1341, doi:10.1016 / j.mri.2012.05.001 (2012)). Radiomic features can be statistics related to shape, first-order statistics, or texture statistics (e.g., second-order and higher-order statistics). Shape-related and first-order radiomic features can be calculated using the original ray density (HU) value of the ROI voxel. For computing texture features (e.g., gray level co-occurrence matrix [GLCM], gray level dependency matrix [GLDM], gray level run length matrix [GLRLM], gray level size zone matrix [GLSZM], and neighborhood gray tone difference matrix [NGTDM], see Tables R1-R7), ROI voxel radiodensity or attenuation values ​​are preferably discretized into multiple bins, preferably 16 bins, preferably with equal widths (e.g., a width of ten HU) to reduce noise while allowing sufficient resolution to detect biologically significant spatial variations in ROI attenuation. Discretization into 16 bins is recommended as the best way to increase the signal-to-noise ratio of radiomics analysis images. However, discretization into more or less than 16 bins is also possible. To enforce symmetric, rotation-invariant results, some or all of the radiomic features, specifically texture statistics (GLCM, etc.), can be computed in all (orthogonal) directions (e.g., four directions on the four sides of a square pixel or voxel) and then averaged (e.g., using the average or other mean of the individually computed values ​​of the feature in each of the directions).

[0100] Some or all of the radiomic features, particularly those related to first-order and texture-based statistics, may also be calculated for a three-dimensional wavelet transform of the raw image data, thereby generating a number of additional radiomic feature sets, for example, as described 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 raw data. At low frequencies (longer time intervals), the wavelets characterize the coarse structure of the data to identify long-term trends. Thus, wavelet analysis allows the extraction of hidden and significant temporal features of the raw data while improving the signal-to-noise ratio of the imaging study. The data can be decomposed into multiple (e.g., eight) wavelet decompositions by discrete wavelet transform by passing the data through a multi-stage (e.g., three-stage) filter bank. At each stage, the data is decomposed into high-frequency components and low-frequency components by high-pass and low-pass filters, respectively. Therefore, if a three-stage filter bank is used, eight wavelet decompositions will be generated, corresponding to HHH, HHL, HLH, HLL, LHH, LHL, LLH, and LLL, where H refers to "high pass" and L refers to "low pass". Of course, more or less than eight levels can be used alternatively to decompose the data. Such decomposition can be performed using widely available software (such as the Slicer radiomics software package incorporating the Pyradiomics library). Optionally, radiomic features can be calculated entirely based on raw (original) data, i.e., wavelet transforms are not applied. Therefore, in the case of a list, group, or cluster of radiomic features disclosed herein, it should be understood that these can be reduced to exclude those radiomic features calculated based on wavelet transforms. Where a radiomic feature is calculated based on a wavelet decomposition or transformation of the data, this is indicated by a suffix indicating the wavelet decomposition that the radiomic feature has been calculated based on (e.g., highpass, highpass, highpassed HHH). Thus, for example, "skewness LLL" indicates the radiomic feature "skewness" calculated based on an LLL wavelet decomposition. In the absence of a suffix, the radiomic feature is calculated based on the raw (or original) data.

[0101] Export radiomics labels

[0102] The present invention provides a method for deriving a radiomics signature for characterizing a ROI (e.g., a region comprising or consisting of adipose tissue or fat), such as for detecting or identifying adipose tissue dysfunction or for predicting metabolic risk, such as the risk of developing a metabolic disorder or disease (e.g., diabetes). The radiomics signature is derived using medical imaging data for a plurality of individuals and data indicating or associated with adipose tissue dysfunction for each of the plurality of individuals (e.g., patients). Specifically, the data indicating adipose tissue dysfunction may include or consist of one or more markers (e.g., biomarkers) of adipose tissue dysfunction, such as the presence or absence of a metabolic disorder, an adipose tissue phenotype associated with adipose tissue dysfunction, or any other information that allows inference as to whether an individual suffers from adipose tissue dysfunction.

[0103] Adipose tissue dysfunction is widely understood to mean the abnormal function of fat that causes, for example, inflammation and fibrosis. As previously mentioned, adipose dysfunction plays a vital role in the development of metabolic disorders such as insulin resistance and diabetes. The data indicating adipose tissue dysfunction can therefore include the information about the presence or absence of metabolic disorders related to adipose tissue dysfunction (such as diabetes or insulin resistance) for each individual in the individual, and the information can be used as the marker of adipose dysfunction. Specifically, data can include the measurement results of one or more features (that is, adipose tissue phenotype) of adipose tissue biology related to adipose tissue dysfunction (such as adipocyte hyperplasia or hypertrophy, inflammation, adipogenesis (adipocyte differentiation) and / or fibrosis) and / or having its characteristic, and these can be used as the marker of adipose tissue dysfunction. Alternatively (or additionally), alternative biomarkers of these features of adipose tissue biology can be used, such as the gene expression profile selected. Specifically, gene expression profiles as markers of characteristics of adipose tissue dysfunction can be used, for example, the data can include gene expression profiles of FABP4 (fatty acid binding protein 4 associated with adipogenesis) and / or TNFA (tumor necrosis factor alpha associated with inflammation). Specifically, as shown in the following example section and Figures 4 (a) and (b), both FABP4 and TNFA gene expression profiles can be used, where high TNFA expression is combined with low FABP4 expression as a surrogate marker for adipose tissue dysfunction and is used to indicate fat dysfunction. In this context, "high" and "low" can be defined relative to the median (or other mean, such as the average) of the expression of the relevant genes, for example, as measured for the multiple individuals or other populations. Therefore, "high" can include values ​​higher than (and optionally also equal to) the median, and "low" can include values ​​lower than (and optionally also equal to) the median. Alternatively, other definitions of "high" and "low" are possible. For example, "high" can mean the highest tertile or quartile, and "low" can mean the lowest tertile or quartile. However, it will typically be the case that the upper limit of "low" will be lower than (or equal to) the mean (e.g., average or median), and the lower limit of "high" will be higher than (or equal to) the mean. Therefore, individuals can be classified, classified or identified as having adipose tissue dysfunction based on their gene expression profiles and individuals with certain gene expression profiles (such as the individuals described above). Other individuals can be classified or identified as not having adipose tissue dysfunction. Based on the markers (e.g., the values ​​of the markers), the multiple individuals can each be identified as having or not having (particularly when collecting medical imaging data) adipose tissue dysfunction.

[0104] A step-by-step approach may be followed to develop a radiomic signature. First, for each of the plurality of individuals, a plurality of radiomic features are calculated from the medical imaging data, for example, as described above. The radiomic features may include some or all of the radiomic features defined in Tables R1-R7, and each of the radiomic features may be calculated based on the original image data and / or one or more wavelet transforms (or wavelet decompositions) of the image data, as described above. Preferably, each of the radiomic features is calculated for the original image data and for the aforementioned eight three-dimensional wavelet decompositions of the image data.

[0105] Unstable features can be removed from the multiple radiomic features. A z-score transformation can be applied to the feature (i.e., the value of the radiomic feature is expressed according to the number of standard deviations from the mean), and a stability analysis can be performed based on the z-score. The stability analysis can include calculating the scan-re-scan stability of the radiomic feature (i.e., stability on multiple imaging data acquisitions or scans) and / or stability on multiple or repeated depictions of the region of interest (e.g., segmentation or contours). Multiple depiction stability analysis can include calculating stability on multiple depictions, each depicted by a different operator or individual. Preferably, the stability analysis includes performing scan-re-scan stability and multiple depiction stability, and only retaining features that are found to be stable according to the two analyses. Stability can be measured by an intraclass correlation coefficient (ICC), and unstable radiomic features can be identified as those with an intraclass correlation coefficient (ICC) in repeated imaging data acquisitions (e.g., imaging scans) and / or in multiple depictions below a stability threshold. For example, the stability threshold can be at least about 0.9, such as about 0.9, so as to exclude all radiomic features with ICC<0.9. However, other stability thresholds, such as 0.85 or 0.95, may be used instead. The scan-re-scan ICC may be calculated for multiple repeated scans, such as two to ten scans, specifically two or ten scans. In other words, stability analysis may be performed on radiomic features removed from multiple radiomic features and unstable radiomic features. The stability analysis may be performed based on the imaging data of the multiple individuals, or may be performed using other data, such as reference data, such as the RIDER dataset ("RIDER: The Reference Image Database to Evaluate Therapy Response"; obtained online from https: / / wiki.cancerimagingarchive.net / display / Public / RIDER+Collections; jsessionid=C78203F71E49C7EA3A43E0D213CE5555).

[0106] Then, the collinearity of the retained radiomic features (i.e., stable radiomic features) can be reduced or eliminated by removing pairwise correlations, i.e., by removing at least one of each pair of identified collinear radiomic features. The removal of pairwise correlations can be performed in a stepwise manner. Collinear radiomic features can be identified as those radiomic features that are correlated to each other to a degree at least equal to a given correlation threshold. The correlation threshold is preferably applicable to positive and negative correlations, for example, the correlation threshold can be expressed as a modulus. Pairwise correlations can be calculated using Spearman's rho coefficient, and the correlation threshold is at least about |rho|=0.75, for example, about |rho|=0.75, so that all pairwise correlations at the level of |rho|≥0.75 are eliminated. Preferably, the correlation threshold can be at least about |rho|=0.9, for example, about |rho|=0.9. As will be readily understood in the art, correlation or collinearity is a measure of how closely two radiomic features change from one individual to the next, and can be calculated based on the measured radiomic feature values ​​of the multiple individuals.

[0107] For example, when a pair of collinear radiomic features are identified, one of the two features is preferably eliminated from the plurality of features. For example, the radiomic features with weaker associations with AT dysfunction calculated from the data can be eliminated, and the radiomic features with the strongest associations with AT dysfunction can be retained, but this is not necessary, and any one can be retained or eliminated. For example, the collinear elimination step can be performed in an unsupervised manner without considering AT dysfunction, and the algorithm can eliminate the most redundant features that have the least effect on the variation of the study population (e.g., features with smaller variances as measured across the plurality of individuals). In one example, when a pair of collinear features are identified, the feature with the largest mean (e.g., average) absolute correlation (i.e., the average correlation value (or average modulus or squared correlation value) with all other radiomic features) is removed. This can be performed in a stepwise manner until no collinear radiomic features remain.

[0108] The collinear elimination step can be performed using an algorithm or function (e.g., function claret::findCorrelation, R software package, see Kuhn, M. & Johnson, K. "Applied Predictive Modelling" (Springer, 2013)). For example, the function or algorithm can construct a pairwise correlation matrix containing pairwise correlations between radiomic features. The function can then search the correlation matrix and return an integer vector corresponding to columns to remove to reduce pairwise correlations. The radiomic features corresponding to these columns can then be removed from the multiple radiomic features. When determining which columns to remove, the algorithm can first identify the pairwise correlations between radiomic features. When two collinear radiomic features are identified, the algorithm then identifies the column corresponding to the feature with the largest mean absolute correlation for removal.

[0109] Regardless of how the collinearity elimination step is performed, the end result is preferably a reduced plurality of radiomic features, wherein each of the features has a correlation with each of the other remaining features that is less than a correlation threshold. In other words, the method may involve the step of removing radiomic features to eliminate collinearity between radiomic features, such that none of the remaining radiomic features is collinear with any of the other remaining radiomic features. This may involve calculating pairwise correlations between radiomic features and removing at least one of any identified collinear feature pairs.

[0110] The reduction of collinearity has the advantage of reducing the redundancy of radiomic features. Collinear features, i.e., those that tend to vary in the same way between individuals, are likely to be sensitive to the same or similar phenotypic aspects of AT. Therefore, the removal of collinear features ensures an increase in the diversity of information in the final signature, with each radiomic feature included in the final signature representing a different phenotypic property of AT.

[0111] A radiomics signature can then be constructed based on at least two of the remaining radiomics features, which survive which steps described above (e.g., stability analysis and / or collinearity elimination) are performed. For example, a radiomics signature can then be constructed based on at least two of the reduced multiple non-collinear radiomics features that survived the collinear elimination step. The reduced multiple features that survived the collinear elimination step are otherwise referred to as "original features." However, since the eliminated radiomics features are each strongly correlated with at least one of the original features, a label in which one or more of the original features are replaced by one or more of the features that are collinear with the replaced original features will generally be performed similarly to a label calculated based only on the original features. For example, one of the original features can be exchanged for one of the features calculated to be collinear with the original feature, and the label should be performed similarly.

[0112] Thus, the process of constructing a radiomic signature may involve constructing "groups" of radiomic features (each group including one of the original features), wherein each of the radiomic features in each group is collinear with at least the "original" feature in the group (i.e., features of the group that survived, for example, a collinear elimination step). In this way, each original feature is associated with its own group of collinear equivalents, which are radiomic features that are calculated to be collinear with the original feature in the data. These groups of collinear equivalents may be constructed for each of the original radiomic features or only for a subset of the original radiomic features, such as those identified, for example, by a feature selection algorithm (see below) to maximize the association of the signature with AT dysfunction. The construction of these groups may be performed in place of the collinear elimination step. For example, instead of eliminating one of each pair of collinear features, the collinear features may be assigned to the same group. Alternatively, a pairwise elimination step may be performed as described above, and then, once the original features are identified, the eliminated features may be reintroduced by assigning them to the group of original features with which they are most correlated or collinear.

[0113] However, no matter how the groups are constructed, the end result should be that each radiomic feature is assigned to the same group or groups as one or more of the original radiomic features with which it is collinear. If a radiomic feature is collinear with two "original" features, it is preferred that the radiomic feature be assigned to the group of the original feature that is most collinear with it, but it is also possible to assign the radiomic feature to the group of all the original features with which it is collinear.

[0114] Therefore, the "original" radiomic feature in each group represents the "partner" radiomic feature of each of the other radiomic features in the group, where each of the radiomic features in each group is collinear with its "partner" feature. Therefore, in this sense, the original radiomic feature can be considered to be its own "partner" radiomic feature because it is completely collinear with itself.

[0115] Although by definition, "raw" features are not collinear with every other feature, some of the raw features will be more similar or internally correlated with each other than other features. Therefore, the "raw" radiomic features themselves can be grouped or "clustered" into multiple clusters of similar or related features. As previously described, the degree of correlation between radiomic features is a measure of the degree to which two radiomic features tend to vary from one another between different individuals. Pairwise correlations can be calculated using Spearman's rho coefficient or other correlation measures (such as Pearson's correlation coefficient). Clustering can be performed, for example, using a hierarchical clustering method (such as a hierarchical clustering algorithm) to classify significant radiomic features into the multiple clusters. Hierarchical clustering can be performed in an unsupervised manner, that is, independent of the strength of the correlation of the radiomic features with AT dysfunction (or its markers). In other words, clustering can be performed on the strength of the correlation of the features with each other, so that the radiomic features are clustered together with those that are most correlated with them. Specifically, the intra-cluster correlation may be stronger than the inter-cluster correlation, that is, the correlation between the features within the cluster is stronger than the correlation between the features in different clusters. The final identification of the clusters can be performed by a person, such as an operator, examining the correlation data. For example, the operator can examine a dendrogram and / or a two-dimensional correlation map (or heat map) representing the hierarchical clustering of radiomic features, which plots the correlation of each radiomic feature with each other radiomic feature (and itself) in the radiomic features, and clusters can be identified based on an inspection of the correlation data within the radiomic features. In the correlation heat map, the radiomic features can be arranged along the x-axis and y-axis of the correlation map in an order determined according to the hierarchical clustering (that is, the features are located near the features with which they are most closely associated or correlated). This visual inspection can be used together with hierarchical clustering to identify the appropriate clustering of radiomic features. Alternatively, clusters can be identified only by hierarchical clustering or visual inspection.

[0116] As identified in the cluster analysis, a radiomic signature can be advantageously constructed based on at least two of the original features selected from different clusters. For example, a radiomic signature can be constructed based on at least one radiomic feature selected from each cluster. If a collinear equivalent of the original feature is included in the signature in place of the original feature, the collinear equivalent preferably corresponds to the original radiomic feature belonging to a different cluster. Preferably, the radiomic signature includes at least two of the original radiomic features (or their collinear equivalents) that are found to maximize the accuracy of the signature to predict a cardiac condition or myocardial disease. Other radiomic features may also be included in the signature, for example, two or more radiomic features from any one or all of the clusters may be included in the signature. However, in order to provide a signature that is more closely associated with AT dysfunction, and therefore has improved diagnostic and prognostic utility, it is preferred to include at least two radiomic features, each from a different cluster. Again, this is because the features from different clusters provide complementary phenotypic information associated with the ROI because the features are weakly associated with each other. Specifically, radiomic features from different clusters may be sensitive to different phenotypic properties of the ROI, because the radiomic features were found not to vary in a similar manner to each other, indicating that the radiomic features are sensitive to different phenotypic properties of the ROI. Therefore, the radiomic signature should preferably be calculated based on (measured values ​​of) at least two radiomic features, each of which is selected from a different cluster. For example, the radiomic signature may include at least three radiomic features, each of which is selected from a different cluster. Preferably, the initial radiomic signature may include one radiomic feature from each cluster.

[0117] A radiomic signature can be constructed based on at least two (or all) of the reduced plurality of features (e.g., "original" features) or their collinear equivalents. For example, if a group of collinear equivalents as described above is constructed, an initial radiomic signature can be constructed based on at least two radiomic features, each radiomic feature being selected from a different group.

[0118] One or more machine learning algorithms can be used to construct a radiomics signature. The machine learning algorithm may include a recursive feature elimination algorithm, such as a random forest algorithm, and may include multiple (e.g., 5-fold) cross validation. Some (preferably all) of the remaining (or "original") radiomics features that survive the collinear elimination step may be input into one or more machine learning algorithms to identify the optimal number and identity of radiomics features that maximize the association of radiomics signatures with AT dysfunction. As mentioned above, collinear equivalents may be used to replace the original radiomics features or in addition to the original radiomics features, and the collinear equivalents may be input into one or more machine learning algorithms. For example, one or more machine learning algorithms may refine the effect of each of the input radiomics features on the label to improve the association of the radiomics label with AT dysfunction (e.g., markers of AT). In doing so, the machine learning algorithm may identify a subset of radiomics features that maximizes the ability of the label to distinguish or identify AT dysfunction, and may calculate the radiomics signature based on at least the subset of radiomics features. In addition to the radiomic features, clinical profile characteristics (e.g., age, sex, diabetes, dyslipidemia, hypertension, HOMA-IR, body mass index, and subcutaneous fat area) can also be input into the machine learning algorithm and can also be optionally included in the final label.

[0119] Specifically, a machine learning algorithm can be used to refine the contribution of radiomic features to the signature to increase or maximize the association of the signature with a cardiac condition or myocardial disease. For example, a plurality (preferably all) of the original radiomic features (i.e., those that survived the collinear elimination step) and / or collinear equivalents of the original radiomic features (e.g., those belonging to the same group) can be input into a first machine learning algorithm. The first machine learning algorithm can be used to identify the optimal number and identity of radiomic features to be included in the signature, specifically to maximize the accuracy of the first machine learning algorithm for distinguishing AT dysfunction. In other words, the construction of the radiomic signature can include a feature selection step, in which the radiomic features to be included in the signature are selected. In the feature selection step, the feature selection algorithm (e.g., a machine learning algorithm) can select a subset of radiomic features to be included in the final signature, specifically a subset of radiomic features that are predicted to maximize the association of the final signature with AT dysfunction. This has the advantage of reducing the complexity of optimizing the final signature because it reduces the number of radiomic features that need to be considered. The first machine learning algorithm that performs this feature selection step can use recursive feature elimination, for example, through a random forest algorithm. However, other algorithms may alternatively be used. The first machine learning algorithm may be constrained to require that the resulting signature includes at least two radiomic features selected from different clusters identified in the cluster analysis step (e.g., using a clustering algorithm). For example, the machine learning algorithm may be constrained to require that the resulting signature includes at least one radiomic feature selected from each of the clusters.

[0120] Once the number and identity of radiomic features to be included in the signature are identified, a second machine learning algorithm can be used to optimize the contribution of each of the features identified by the first machine learning algorithm to the signature. In other words, the construction of the radiomic signature can include the step of using a second machine learning algorithm to refine or optimize the radiomic signature to increase or maximize the association of the radiomic signature with AT dysfunction. The second machine learning algorithm can be a logistic regression algorithm. The data for the training cohort can be used to derive or refine the signature, and the data for the test cohort can be used to validate the signature. For example, an internal cross-validation can be used to construct the signature. The internal cross-validation can be multiple, for example, 5 times.

[0121] Alternatively, a single machine learning algorithm can be used, rather than in the two-step process described above, to construct a signature based on radiomic features, e.g., based on "raw" radiomic features. For example, a single machine learning algorithm can be used to identify features to be included in the final signature, and to refine or optimize the signature. For example, the number of radiomic features to be included in the final signature can be pre-selected or predetermined and input as a parameter to the machine learning algorithm, or the number can be left open and selected by the machine learning algorithm itself, which can also refine or optimize the radiomic signature. In other words, the machine learning algorithm can contain feature selection functionality.

[0122] Typically, the construction of a radiomic signature involves refining or optimizing the radiomic signature, for example, each effect of each radiomic feature in the radiomic signature on the signature, to improve the correlation or association of the signature with AT dysfunction based on the data. For example, the signature may include a weighted sum of the values ​​of each radiomic feature in the radiomic signature contained in the initial signature, and the weights of each radiomic feature in the radiomic signature may be gradually optimized or refined. The coefficient by which each radiomic feature in the radiomic signature is multiplied is generally referred to as a beta (β) coefficient, and it is these β coefficients that may be optimized or refined. As will be readily appreciated by those skilled in the art, the association of the radiomic signature with AT dysfunction may be calculated based on a receiver operating characteristic (ROC) curve analysis, specifically using an area under the curve (AUC) measurement (i.e., a C statistic).

[0123] Preferably, the raw features that survived the collinearity elimination step are input into one or more machine learning algorithms, and the one or more machine learning algorithms identify a subset of the raw features that maximize the accuracy of the signature for classification of AT dysfunction. A radiomic signature is then preferably constructed from two or more radiomic features that belong to a group of collinear equivalents corresponding to the raw radiomic features, wherein the two or more radiomic features are selected from different groups.

[0124] This method is advantageous because in order to provide a signature that is more closely associated with AT dysfunction, and therefore has improved diagnostic and prognostic utility, the signature preferably comprises at least two radiomic features, each from a different group. This is because features from different groups (and therefore corresponding to different non-collinear original features) provide complementary information related to the ROI. Specifically, radiomic features from different groups will be sensitive to different phenotypic properties of the ROI, and the radiomic features are collinear with different "original" or "partner" features. For example, the starting radiomic signature may include at least three radiomic features, each of which is selected from a different group. Alternatively, the starting radiomic signature may include at least four radiomic features, each of which is selected from a different group. Alternatively, the starting radiomic signature may include at least five radiomic features, each of which is selected from a different group. Preferably, the initial radiomic signature may include one radiomic feature from each group.

[0125] As mentioned above, a label may include a weighted sum of the calculated values ​​of multiple radiomic features. Labels may also contain other terms, such as addition or subtraction of constants or multiplication factors. However, typically, the label will be linearly related to the weighted sum of the radiomic feature values ​​in some way.

[0126] The radiomics signature may take the form of or include the terms described (eg, the signature may be calculated based on a function including the terms):

[0127] Σb i rf i +c

[0128] Where c is a constant (can be zero or non-zero), b i is the weighting coefficient (or β parameter) of radiomic feature i, and rf i is the measured value of radiomic feature i. The constant c is not required but can be included to ensure that all result values ​​are positive or negative.

[0129] Radiomics signature

[0130] The radiomic signature of the present invention is calculated based on the measured values ​​of the radiomic features obtained from the medical imaging data. In particular, the radiomic signature is preferably calculated based on at least two radiomic features.

[0131] As described above, in order to improve the prognostic and diagnostic value of the signature, the signature is preferably calculated based on at least two different radiomic features selected from different clusters of similar or related raw radiomic features. This reduces redundancy and improves the diversity of information contained in the signature calculation, because the features from different clusters are related to different texture aspects of the ROI.

[0132] Three clusters (AC) of "raw" (i.e., non-collinear) radiomic features have been identified using a hierarchical clustering algorithm (see Examples and Figure 3(c)). The members of the three clusters are identified in Table 1 (and Figure 3(c)). The radiomic signature may include at least two of the non-collinear radiomic features from Table 1. Advantageously, the radiomic signature may be calculated based on radiomic features of at least two clusters selected from clusters AC identified in Table 1, at least two of the radiomic features being selected from different clusters. Preferably, the radiomic signature may be calculated based on at least one radiomic feature selected from each of the clusters AC identified in Table 1.

[0133] Table 1: Clustering of radiomic features Clustering of radiomic features

[0134]

[0135]

[0136] Five radiomic features were found to maximize the association of the radiomic signature with AT dysfunction (see examples below), and these are listed in Table 2. The radiomic signature can advantageously be calculated based on at least two of the radiomic features from Table 2. Preferably, the radiomic signature is calculated based on at least two radiomic features from Table 2, each of the at least two radiomic features being selected from a different cluster. Further preferably, the radiomic signature can be calculated based on at least one radiomic feature selected from each of the clusters AC identified in Table 2. In order to maximize the association of the radiomic signature with AT dysfunction, the radiomic signature is preferably calculated based on all of the radiomic features listed in Table 2.

[0137] Table 2: Further optimized radiomics feature clustering

[0138]

[0139] As previously mentioned, the "original" radiomic features of Tables 1 and 2 can be replaced by other radiomic features that are collinear with the substituted "original" radiomic features (i.e., collinear equivalents) to obtain a signature with similar diagnostic and prognostic utility. Therefore, a radiomic signature can be calculated based on (i.e., including) at least two radiomic features selected from the radiomic features of Table 3. Each of the groups identified in Table 3 contains one of the five "original" radiomic features listed in Table 2, which have been found to maximize the association of the signature with AT dysfunction and those radiomic features, which have been calculated to be collinear with the original radiomic features to a degree of at least |rho|=0.75, where rho is Spearman's rho. Therefore, a radiomic signature can be constructed as described above, but one or more of the radiomic features of the radiomic features of Table 2 are replaced with radiomic features that are collinear with the features, as described in Table 3. For example, a radiomics signature can be calculated based on at least two radiomics features selected from different groups of Table 3. For example, a radiomics signature can be calculated based on at least three radiomics features selected from different groups of Table 3. For example, a radiomics signature can be calculated based on at least four radiomics features selected from different groups of Table 3. For example, a radiomics signature can be calculated based on at least five radiomics features selected from different groups of Table 3. Specifically, a radiomics signature can be calculated based on at least two radiomics features selected from a group corresponding to original features belonging to different clusters AC. Specifically, a radiomics signature can be calculated based on at least one radiomics feature selected from each cluster of clusters AC identified in Table 3 below.

[0140] Table 3. Radiomics feature groups co-linear with the radiomics features in Table 2

[0141]

[0142]

[0143]

[0144]

[0145]

[0146]

[0147] The groups identified in Table 3 can be simplified to include only those radiomic features that have a correlation of at least |rho|=0.800 with the original feature of the group (i.e., one of the five original features identified by the feature selection algorithm) (this includes the original feature itself, which, by definition, has a correlation of rho=1 with itself). For example, the groups can be simplified to those features that have a correlation of at least |rho|=0.850 with the original features of the group. For example, the groups can be simplified to those features that have a correlation of at least |rho|=0.900 with the original features of the group. For example, the groups can be simplified to those features that have a correlation of at least |rho|=0.950 with the original features of the group.

[0148] In addition to calculating a radiomic signature based on at least two radiomic features from different clusters or groups, a radiomic signature can also be calculated based on additional radiomic features. For example, a radiomic signature can include more than one radiomic feature from any given cluster or group, or can include radiomic features that are not included in any cluster or group in the cluster or group. Therefore, it can be said that a radiomic signature is calculated based on multiple radiomic features, and the multiple radiomic features can include at least two radiomic features mentioned above.

[0149] Each of the radiomic signatures of the present invention provides a direct means for characterizing ROIs and identifying AT dysfunction using medical imaging data. Because each of the radiomic signatures of the present invention is based on a relatively small number of the total number of possible radiomic signatures that can be measured, the signature is easy to calculate and understand, and clinicians can better understand its physiological significance.

[0150] system

[0151] The method of the present invention can be performed on a system such as a computer system. Therefore, the present invention also provides a system, which is configured or arranged to perform one or more of the methods of the present invention. For example, the system may include a computer processor, which is configured to perform one or more of the methods or steps of the present invention. The system may also include a computer readable memory loaded with executable instructions for performing the steps of any of the methods of the present invention.

[0152] Specifically, the method of deriving radiomics labels can be performed on such a system, and therefore such a system is provided according to the present invention. For example, the system can be configured to receive and optionally store a data set, the data set including the values ​​of multiple radiomics features of ROI obtained from medical imaging data for each individual in a plurality of individuals. The system can be configured to use such a data set to construct (e.g., derive and verify) a radiomics label according to the method of the present invention.

[0153] Alternatively, the system can be configured to perform a method of characterizing an ROI or identifying AT dysfunction. Specifically, the present invention provides a system for characterizing an ROI using medical imaging data of a subject. The system can be configured to calculate a value of a radiomics signature of the ROI using the medical imaging data. The radiomics signature can be calculated based on the measured values ​​of at least two radiomics features of the ROI, and the measured values ​​of at least two radiomics features can be calculated based on the medical imaging data.

[0154] The system may also be configured to compute radiomic features from medical imaging data, as described in more detail above.Thus, the system may be configured to receive and optionally store medical imaging data, and process the imaging data to compute radiomic features.

[0155] Definition of radiomic features

[0156] The definitions of radiomic features referred to herein are generally well understood in the radiomics field simply by referencing the name of the radiomic feature. However, for ease of reference, the definitions of the features used herein are provided in Tables R1 to R7 herein. The radiomic features in Tables R1 to R7 are defined based on the radiomic features used by the Pyradiomics software package (http: / / pyradiomics.readthedocs.io / en / latest / features.html, see van Griethuysen, JJM, Fedorov, A., Parmar, C., Hosny, A., Aucoin, N., Narayan, V., Beets-Tan, RGH, Fillon-Robin, JC, Pieper, S., Aerts, HJWL (2017) "Computational Radiomics System to Decode the Radiographic Phenotype." Cancer Research, 77(21), e104–e107. https: / / doi.org / 10.1158 / 0008-5472.CAN-17-0339). Most of the features defined in Tables R1 to R7 conform to the feature definitions described by the Imaging Biomarker Standardization Initiative (IBSI), which are available in: Zwanenburg et al. (2016) (Zwanenburg, A., Leger, S., Vallières, M. and S. (2016) "Image biomarker standardisation initiative - feature definitions" arXiv electronic preprint: 1612.07003 [cs.CV]). Where the definitions provided below do not fully conform to the IBSI definitions, it will be appreciated that either definition may be used in accordance with the present invention. Ultimately, the precise mathematical definition of radiomic features is not critical, as slight modifications will not affect the general properties of the image measured by each of the features. The present invention is therefore intended to encompass slight modifications to features (e.g., addition or subtraction of constants or scaling) as well as alternative definitions of features.

[0157] a. First-order statistics

[0158] These statistics describe the central tendency, variability, uniformity, asymmetry, skewness, and size of the attenuation values ​​in a given region of interest (ROI), while ignoring the spatial relationship of individual voxels. In this way, these statistics describe the quantitative and qualitative characteristics of the entire ROI. For each of the eight wavelet transforms and the original CT image, a total of 19 features were calculated, as follows:

[0159] make:

[0160] ●X is a set of N contained in the region of interest (ROI) p Attenuation or radiodensity value of the voxel (e.g., in HU)

[0161] ●P(i) is a g The first-order histogram of discrete intensity levels, where N g is the number of non-zero bins, equal to the width of the zero interval.

[0162] ●P(i) is the normalized first-order histogram and is equal to

[0163] c is the value by which the intensity is biased to prevent negative values ​​in X. This ensures that the voxels with the lowest grayscale values ​​contribute the least to the energy, rather than the voxels with grayscale intensity closest to 0. Since the HU range of adipose tissue (AT) within the ROI (-190 to -30 HU) does not include zero, c can be set to c = 0. Therefore, higher energies correspond to less radiodense AT and therefore higher lipophilic content.

[0164] ●∈ is an arbitrarily small positive number (e.g., ≈2.2×10 -16 )

[0165] Table R1: First-order radiomic features used for ROI characterization

[0166]

[0167]

[0168] b. Shape-related statistics

[0169] Shape-related statistics describe the size and shape of a given ROI, regardless of the attenuation values ​​of its voxels. Since shape-related statistics are independent of grayscale intensity, they are consistent across all wavelet-transformed and raw CT images and are therefore calculated only once. These definitions are as follows:

[0170] make:

[0171] V is the volume of ROI, in mm 3 Unit

[0172] A is the surface area of ​​ROI, in mm 2 Unit

[0173] Table R2: Shape-related radiomic features for ROI characterization

[0174]

[0175]

[0176] c. Gray Level Co-occurrence Matrix (GLCM)

[0177] In simple terms, the GLCM describes the number of times a voxel with a given attenuation value i is located next to a voxel with j. g ×N g The GLCM describes the second-order joint probability function of the image region constrained by the mask and is defined as P(i,j|δ,θ). The (i,j)th element of this matrix represents the number of times the combination of levels i and j appears 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 according to the infinite norm. For δ=1, this results in 2 neighbors for each of the 13 angles in 3D (26-connectivity), and for δ=2, 98-connectivity (49 unique angles). In order to obtain rotationally invariant results, the statistics are calculated in all directions and then averaged to ensure a symmetric GLCM.

[0178] make:

[0179] ∈ is an arbitrarily small positive number (for example, ≈2.2×10 -16 )

[0180] P(i,j) is the co-occurrence matrix of any δ and θ

[0181] p(i,j) is the normalized co-occurrence matrix and is equal to

[0182] Ng is the number of discrete intensity levels in the image

[0183] is the marginal row probability

[0184] is the marginal column probability

[0185] μ x For p x The average gray level intensity is defined as

[0186] μ y for p y The average gray level intensity is defined as

[0187] σ x For p x Standard Deviation

[0188] σ y For p y Standard Deviation

[0189] Where i+j=k, and k=2,3,…,2N g

[0190] where |ij|=k, and k=0,1,…,N g -1

[0191] For p x Entropy

[0192] For p y Entropy

[0193]

[0194] For distance weighting, the GLCM matrix is ​​weighted by a weighting factor W and then summed and normalized. The weighting factor W for the distance between neighboring voxels is calculated by: where d is the distance of the associated angle.

[0195] Table R3: Gray-level co-occurrence matrix (GLCM) statistics for ROI characterization

[0196]

[0197]

[0198]

[0199] d. Gray Level Size Zone Matrix (GLSZM)

[0200] Gray-Level Size Zones (GLSZM) describes the gray-level zones in a ROI, defined as the number of connected voxels sharing the same gray-level intensity. A voxel is considered connected if the distance according to the infinity norm is 1 (26-connected zones in 3D, 8-connected zones in 2D). In the 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 rotation-independent, only one matrix is ​​calculated for all orientations in the ROI.

[0201] make:

[0202] N g is the number of discreet intensity values ​​in the image

[0203] N s is the number of cautious area sizes in the image

[0204] N p is the number of voxels in the image

[0205] N z is the number of regions in the ROI, which is equal to and

[0206] 1≤N z ≤N p

[0207] P(i,j) is the size region matrix

[0208] p(i,j) is the normalized size region matrix, which is defined as

[0209] ∈ is an arbitrarily small positive number (for example, ≈2.2×10 -16 ).

[0210] Table R4: Gray Level Size Zone Matrix (GLSZM) Statistics for ROI Characterization

[0211]

[0212]

[0213]

[0214] e. Gray Level Run Length Matrix (GLRLM)

[0215] The grayscale run length matrix (GLRLM) describes the grayscale runs of consecutive pixels with the same grayscale value, where the grayscale run is defined as the length in number of pixels. In the grayscale run length matrix P(i,j|θ), the (i,j)th element describes the number of runs of grayscale level i and length j that appear in the image (ROI) along angle θ.

[0216] make:

[0217] N g is the number of discreet intensity values ​​in the image

[0218] N r is the number of cautious run lengths in the image

[0219] N p is the number of voxels in the image

[0220] N z (θ) is the number of runs in the image along angle θ, which is equal to And 1≤N z (θ)≤N p

[0221] P(i,j|θ) is the run length matrix of any direction θ

[0222] p(i,j|θ) is the normalized run length matrix, which is defined as

[0223] ∈ is an arbitrarily small positive number (for example, ≈2.2×10 -16 ).

[0224] By default, the value of the feature is calculated separately for each angle on the GLRLM, and the average of these values ​​is then returned. If distance weighting is enabled, the GLRLM is weighted by the distance between neighbor voxels, then summed and normalized. The features are then calculated on the resulting matrix. The distances between neighbor voxels are calculated for each angle using the norm specified in "weightingNorm" Table R5: Grayscale run length matrix (GLRLM) statistics for ROI characterization

[0225]

[0226]

[0227]

[0228] f. Neighborhood Grayscale Tone Difference Matrix (NGTDM) features

[0229] The neighborhood grayscale tone difference matrix quantifies the difference between the grayscale value and the average grayscale value of its neighborhood within the distance δ. The sum of the absolute differences of grayscale level i is stored in the matrix. gl is a set of segmented voxels and x gl (j x ,j y ,j z )∈X gl is the position (j x ,j y ,j z ), then the average gray level of the neighborhood is:

[0230]

[0231] Where (k x ,k y ,k z )≠(0,0,0) and x gl(j x +k x ,j y +k y ,j z +k z )∈X gl

[0232] Here, W is the number of voxels in the neighborhood that are also located in X gl middle.

[0233] make:

[0234] n i is X with gray level i gl The number of voxels in

[0235] Nv,p is X gl The total number of voxels in the i (i.e., the number of voxels with a valid region; at least 1 neighbor). N v,p ≤N P , where N p is the total number of voxels in the ROI.

[0236] p i is the gray level probability and is equal to n i / N v

[0237] is the sum of absolute differences of gray level i

[0238] N g The number of gray levels for caution

[0239] N g,p is the number of gray levels, where p i ≠0

[0240] Table R6: Neighborhood Gray Tone Difference Matrix (NGTDM) for ROI characterization

[0241]

[0242]

[0243] g. Gray Level Dependence Matrix (GLDM)

[0244] The gray-level dependency matrix (GLDM) quantifies the gray-level dependency in an image. Gray-level dependency is defined as the number of connected voxels within a distance δ that depend on the central voxel. A neighborhood voxel of gray level j is considered to depend on the central voxel of gray level i if |ij|≤α. In the gray-level dependency matrix P(i,j), the (i,j)th element describes the number of times a voxel of gray level i with j dependent voxels in the neighborhood appears in the image.

[0245] N g is the number of discreet intensity values ​​in the image

[0246] N d The number of cautiously dependent dimensions in the image

[0247] N z is the number of dependency regions in the image, which is equal to

[0248] P(i,j) is the dependency matrix

[0249] p(i,j) is the normalized dependency matrix, which is defined as Table R7: Gray Level Dependence Matrix (GLDM) statistics for ROI characterization

[0250]

[0251]

[0252]

[0253] Examples

[0254] method

[0255] Study Design

[0256] Study group 1 consisted of 225 patients from the Oxford Heart Vascular and Fatty (OxHVF) cohort undergoing coronary artery bypass grafting (CABG). Exclusion criteria were any inflammatory, infectious liver / renal disease or malignancy. Patients receiving nonsteroidal anti-inflammatory drugs were also excluded. All patients underwent non-contrast enhanced CT scans, as described below, with the aim of exploring whether non-invasive radiomic phenotyping of adipose tissue (AT) could provide information about AT biology and metabolic risk. In group A (n=167), adipose tissue samples were collected from subcutaneous fat (from a chest incision) during surgery. Samples were snap-frozen for gene expression studies, as described below.

[0257] Study group 2 consisted of 40 patients who underwent 18-fluorodeoxyglucose positron emission tomography / computed tomography (PET / CT) imaging for clinical indications. This group was used to externally validate the selected radiomics signature as the gold standard imaging modality for quantification of tissue inflammation in vivo (i.e. 18F FDG PET / CT) as a biomarker of adipose tissue inflammation.

[0258] Gene expression studies

[0259] Samples of adipose (or adipocytes) and aortic tissue were snap frozen in QIAzol (Qiagen) and stored at -80°C. RNA was extracted using RNeasy Micro or Mini Kit (Qiagen) and ribonucleic acid was converted to complementary DNA (Quantitect Rev. Transcription Kit-Qiagen). The cDNA was then subjected to quantitative polymerase chain reaction (qPCR) using TaqMan probes (Applied Biosystems) for FABP4 (Assay ID: Hs00609791_m1), PREF1 (Assay ID: Hs00171584_m1), TNFA (Assay ID Hs01113624_g1), IL6 (Assay ID Hs00985639_m1), Col1A1 (Assay ID Hs00164004_m1) and Sfrp5 (Assay ID Hs00169366_m1). Cyclophilin A (PPIA) was used as a housekeeping gene (assay ID Hs04194521_s1). On an ABI 7900HT fast real-time PCR system (Applied Biosystems), 5 ng cDNA was used per reaction and the reactions were performed in triplicate in 384-well plates. The reaction efficiency in each plate was determined based on the slope of the standard curve; the expression of each gene of interest relative to its housekeeping gene was calculated using the Pfaffl method.

[0260] Computerized tomography studies

[0261] Participants in study groups 1 and 2 underwent CT imaging using a 64-slice scanner (LightSpeed ​​Ultra, General Electric). Heart rate was optimized using intravenous injections of beta blockers, and sublingual glyceryl trinitrate (800 μg) was also administered to achieve maximal coronary vasodilation. Non-contrast, prospective ECG-triggered axial acquisition CT scans (rotation time of 0.35 seconds, axial slice thickness of 2.5 mm, detector coverage of 20 mm, tube energy of 120 kV and 200 mA) were obtained using the carina and diaphragm as cranial and coccygeal landmarks, respectively. The lung field of view was extended to cover the entire chest soft tissue (for adipose tissue analysis). To analyze the adipose tissue radiomics features, the original dicom images were transferred to a dedicated workstation and further post-processed using 3D Slicer (see below).

[0262] Coronary artery calcium score

[0263] By calculating the Agatston score, in Aquarius Coronary artery calcium scores (CCS) were measured in all coronary arteries (RCA alone).

[0264] Radiomic feature extraction of adipose tissue

[0265] Calculation of radiomic features in subcutaneous adipose tissue was performed in CT scans using 3D Slicer software (v.4.9.0-2017-12-18r26813, available at http: / / www.slicer.org; Fedorov, A. et al. "3D Slicer as an image computing platform for quantitative imaging networks" Magnetic Resonance Imaging 30, 1323-1341, doi: 10.1016 / j.mri.2012.05.001 (2012)). Segmentation of adipose tissue was performed by selecting 2.5 mm thick axial slices at the level of the caudal sternum. All voxels located outside the sternocostal region were tracked by manual contouring and identification of fat using a -190 to -30 Hounsfield unit (HU) mask. The segmented adipose tissue was then used to calculate and extract a series of radiomic features using the SlicerRadiomics extension of 3D Slicer, which incorporates the Pyradiomics library into 3D Slicer (see van Griethuysen, JJM et al. "A computational radiomics system for decoding radiographic phenotypes" Cancer Research 77, e104-e107, doi: 10.1158 / 0008-5472.CAN-17-0339 (2017)). The original HU values ​​of the segmented adipose tissue were used to calculate shape-related and first-order radiomic features. In order to calculate texture features (GLCM, GLDM, GLRLM, GLSZM, and NGTDM), the AT voxels were discretized into 16 boxes of equal width (width of 10 HU) to reduce noise while allowing sufficient resolution to detect biologically significant spatial variations in adipose tissue attenuation. In order to strengthen the symmetric, rotation-invariant results, texture statistics (GLCM, etc.) were calculated in all four directions and then averaged.

[0266] Wavelet Transform: First-order and texture-based statistics were also calculated for the three-dimensional wavelet transform of the original images, resulting in eight additional sets of radiomic features.

[0267] Positron Emission Tomography / Computed Tomography Imaging Studies

[0268] Using PET / CT, the gold standard method for assessing tissue inflammation in vivo, the radiomics signature of selected adipose tissue was targeted. 18F Verification of FDG uptake was performed as clinically indicated. 18F Paired PET and CT images of 40 subjects (25 men / 15 women) scanned with FDG PET / CT 18FFDG uptake. PET / CT examinations were performed on a 3D mode time-of-flight (ToF) GE Discovery 690 PET / CT system (GE Healthcare). Patients fasted for at least 6 hours before their scans. 4 MBq / kg of 18F FDG measures the patient's blood glucose. Imaging began 90 minutes (93 ± 7 minutes) after injection and covered the skull base to the upper thigh. PET / CT images were collected under normal tidal breathing for 4 minutes in each bed. CT was performed using 0.984, 120kV, automatic mA pitch and a noise index of 25. PET images were reconstructed using two different algorithms, both of which used CT for attenuation correction and the same normalization correction factor. The standard of care PET reconstruction algorithm used was ToF OSEM (VPFX, GE Medical). This was used with two iterations, 24 subsets and a 6.4mm Gaussian filter. The sinusoids generated at the time of scanning were retrospectively processed using a new ToF BPL reconstruction algorithm (Q.Clear, GE Medical). The acquired PET images were then analyzed using Terarecon Aquarius iNtuition V.4.4.11 software. A circular region of interest (ROI) was drawn at the midline in front of the sternum to calculate the average SUV in subcutaneous adipose tissue. To calculate the target to background ratio (TBR), the mean SUV of adipose tissue was normalized to the mean SUV of the pulmonary artery lumen (at the level of the PA bifurcation).

[0269] Statistical analysis

[0270] Principal component analysis: In panel 1, all 843 computed AT radiomics features were included in principal component analysis to identify principal components (PCs) describing most of the phenotypic variation in the study population. Scrub plots of the PCs against their eigenvalues ​​were constructed. Associations with demographic, biochemical, and adipose tissue gene expression data were explored in relevant correlation plots using the three first components (PC1, PC2, PC3).

[0271] Feature selection and stability assessment: In order to limit the analysis to the most valuable radiomic features as imaging biomarkers, stability assessment was performed on all 843 different radiomic features. For this purpose, 30 paired scans from the RIDER data set (RIDER: a reference image database for evaluating therapy response; obtained online from https: / / wiki.cancerimagingarchive.net / display / Public / RIDER+Collections; jsessionid=C78203F71E49C7EA3A43E0D213CE5555) were used to assess the scan-re-scan ICC of each radiomic feature. Then, in 20 patients from the OxHVF cohort (depicted by two independent operators), the ICC of all radiomic features was also multiplexed for the region of interest. In additional analysis, only those radiomic features (n=196) with scan-re-scan and multiple delineation intraclass correlation coefficients (ICC) ≥ 0.90 were included.

[0272] Unsupervised clustering of the study population by adipose tissue radiomics features: 196 selected radiomics features of AT in panel 1 were converted to Z scores for further analysis. Hierarchical clustering was then performed on the observations in panel 1 using all 196 radiomics features (using the Ward method and squared Euclidean distance, hclust R package). In the n = 225 observations of the panel 1 cohort, the changes in each of the 196 different radiomics features were represented in a correlation heat map, with row dendrograms indicating the clustering of patients. Differences in the distribution of risk factors, biochemical or AT gene expression data between the two clusters of patients were then explored, as appropriate, by using chi-square or independent t tests.

[0273] Feature selection and relevance to AT biology: The relationship between 196 stable radiomic features was visually examined in the similarity-dissimilarity diagram. Next, a stepwise approach was applied, and highly correlated radiomic features (|rho|>0.9) were removed from other analyses by applying a known automated algorithm (see Kuhn, M. & Johnson, K. "Applied Predictive Modeling" (Springer, 2013)). This function (function caret::findCorrelation, R package) searches the correlation matrix and returns an integer vector corresponding to the column to remove in order to reduce paired correlations. Consider the absolute value of paired correlations. If two variables have a high correlation, the function looks at the average absolute correlation of each variable and removes the variable with the maximum average absolute correlation.

[0274] Specific radiomic features were then identified as imaging biomarkers describing adipose tissue biology. The final 32 radiomic features (listed in Table 1) that survived the steps described above were included in the logistic regression model to find independent associations with the expression levels of target genes (TNFA, FABP4) in human AT, which were used as surrogates for the features of adipose tissue biology (inflammation, adipocyte differentiation, fibrosis, adipogenic capacity, and adipokine expression). The expression levels of the target genes were divided into multiple groups (based on the highest tertile) and used as categorical variables in the multivariate logistic regression model. With the target gene of interest as the dependent variable, the final 31 radiomic features (after testing for collinearity by the variance inflation factor, 1 feature was excluded from the model) were fed into the logistic regression model. The classical risk factors (age, sex, diabetes, hypertension, smoking), HOMA-IR, body mass index, and subcutaneous fat area of ​​the baseline model were adjusted (model 1); the radiomic features selected by backward selection (input level: 0.05, removal level: 0.10) of model 2 were further adjusted. The discrimination values ​​of the nested models for identifying AT biosignatures were compared with the help of the corresponding C-statistics and the change in area under the curve (Δ[AUC]).

[0275] Machine Learning and Internal Cross-Validation: Machine learning was used for internal cross-validation of the filtered radiomic features and final feature selection. Machine learning with a random forest algorithm (rfecontrol, caret package R) was used by 5-fold cross-validation to select the main features capable of classifying dysfunctional adipose tissue (i.e., high adipose tissue inflammation plus low adipocyte differentiation, based on the median TNFA and FABP4 gene expression levels, respectively). The top five radiomic features identified from this algorithm validation process (mean LLL, total energy HHL, large dependency low gray level focus, dependency entropy, and run entropy) were used to calculate a composite radiomic signature (referred to as FatHealth) as a measure of adipose tissue health (FatHealth = Σb i x i +c, where c is a constant, x i is the measured value of radiomic feature i, and b i is the beta coefficient of feature i derived from the logistic regression model using the five radiomic features as independent predictors of adipose tissue health. Continuous variables between the two groups were compared by Student's t-test, while categorical variables were compared using Pearson's chi-square test. Analyses were performed using R v3.1.4 (packages: caret, hclust) and SPSS version 25.0. Unless otherwise stated, all tests were two-sided and α was set to 0.05.

[0276] result

[0277] Radiomics feature extraction and principal component analysis

[0278] The study design is outlined in Figure 1 CT scans of 225 patients undergoing coronary artery bypass grafting were used to compute a total of 843 radiomic features through segmentation of subcutaneous adipose tissue. These included 15 shape-related features, 18 first-order statistics, 15 gray-level co-occurrence matrices (GLCMs), 18 gray-level dependency matrices (GLDMs), 16 gray-level run-length matrices (GLRLMs), 16 gray-level size zone matrices (GLSZMs), and 5 neighborhood gray-level hue difference matrices (NGTDM) features, as well as eight wavelet transforms of each of these, as defined in Tables R1 to R7. Initially, exploratory data analysis was performed by reducing the raw radiomics dataset with potentially correlated features to its principal components. A total of 86 components accounted for 99.5% of the variation in the study population (scrunch plot, Figure 2(a)), while the first 3 components explained 60.6% of the observed variation (Figure 2(b)). The first three components were variably correlated with clinical demographics, biochemical characteristics, and gene expression patterns of subcutaneous adipose tissue of group 1 patients (Figure 2(c)), indicating that standard non-contrast enhanced CT images of human adipose contain rich extractable information that is associated with different biological phenotypes and metabolic risk characteristics of adipose tissue.

[0279] Unsupervised clustering based on adipose tissue radiomic phenotypes

[0280] Since principal components are intrinsic to the sample population studied and not transferable values ​​as quantifiable biomarkers, analyses were performed on the radiomic features themselves. Based on the initial library of 843 measured radiomic features, a stability assessment was performed and scan-rescan (RIDER dataset) and multiple profiles (OxHVF cohort, Figure 1 ). Only those radiomic features with ICC ≥ 0.90 (n = 196) were included in further analysis. This set of 196 radiomic features was then used to perform unsupervised hierarchical clustering on the population in panel 1 (Fig. 3a). Hierarchical clustering based on adipose tissue radiomic features identified two different patient clusters that differed significantly in the incidence of metabolic risk factors and insulin resistance states (Fig. 3(a)). These findings suggest that radiomic features of fat depots (e.g., subcutaneous fat depots) should be useful for non-invasive phenotyping and metabolic risk assessment of their biology.

[0281] Incremental value of radiomic signatures for detecting adipose tissue dysfunction

[0282] After showing the proof of concept of the radiomics characteristics of adipose tissue and its biological connection, the radiomics labels that can be used as metabolic risk biomarkers were identified next. The construction of similarity-non-similarity diagram (Fig. 3 (b)) shows that adipose tissue radiomics characteristics are gathered in different groups with highly similar intra-clustering features. These 196 stable radiomics features are variably related to the steady-state model assessment of insulin resistance (HOMA-IR). In order to ensure the information diversity of a higher amount in the final label, an automatic algorithm was applied, which removed highly correlated features in a step-by-step manner. The filtered 32 radiomics features of the final group (listed in Table 1), their mutuality relationship and clustering are presented on the correlation matrix (Fig. 3 (b)).

[0283] Adipocyte differentiation state and inflammation of adipose tissue are the main features of adipose tissue dysfunction and the development of insulin resistance. We first explored the correlation between the gene expression profiles of TNFA and FABP4 in subcutaneous adipose tissue and clinical phenotypes. Patients were divided into four subgroups based on the median levels of TNFA and FABP4. The subgroup with low FABP4 and high TNFA expression had a higher prevalence of diabetes in non-diabetic patients ( Figure 4a ) and insulin resistance ( Figure 4(b) ). Therefore, we next explored whether the radiomic signature of adipose tissue could be used to capture adipose tissue biology, namely its differentiation and inflammatory states.

[0284] To this end, the radiomic signature of subcutaneous adipose tissue was included in a multivariate logistic regression model by a backward selection process using the highest tertile of FABP4 and TNFA gene expression as dependent variables, respectively. Inclusion of the radiomic signature in the multivariate model significantly improved the discrimination of the baseline model for adipocyte differentiation (Δ[AUC]=0.121, p=0.0086, Figure 4(c)) and adipose tissue inflammation (Δ[AUC]=0.287, p=0.0001, Figure 4(d)) status above and below clinical risk factors, HOMA-IR or obesity measures.

[0285] Internal validation and construction of a radiotranscriptomic score for metabolic risk

[0286] The filtered radiomic features were then fed into a machine learning algorithm along with clinical profile properties (age, sex, diabetes, dyslipidemia, hypertension, HOMA-IR, body mass index, and subcutaneous fat area) to identify the main radiomic features that can capture adipose tissue dysfunction (defined as high TNFA and low FABP4 expression based on their median values). A recursive feature elimination algorithm with five-fold cross-validation identified a set of seven variables that maximized the accuracy of the adipose tissue dysfunction model (Figure 4(e)). These main predictors included age and body mass index and a set of five radiomic features: large dependence low gray level emphasis (LDLGLE, GLDM texture feature, used to measure the joint distribution of large dependence with lower gray level values), mean LLL (mean decay calculated on the LLL wavelet transform), run entropy (RE, GLRM feature and heterogeneity index in adipose tissue texture pattern), dependence entropy (DE, GLDM feature used to quantify gray level dependence in images), and total energy HHL (first-order feature, where the lowest gray value has the least contribution calculated on the HHL wavelet transform). The combination of these five radiomic features (listed in Table 2) increased the c-statistic of the baseline clinical model for identifying dysfunctional adipose tissue ((Δ[AUC]=0.124, p=0.001, Figure 4(f)). The exact form of the radiomic signature (FatHealth) is as follows:

[0287] Tag(FatHealth)=

[0288] ((((-1264629630*(large dependency low gray level emphasis)

[0289] +(259259.2593*average_LLL)

[0290] +(101629629.6*dependency entropy)

[0291] +(-121148148.1*Run Entropy)

[0292] +(total energy_HHL)+372222222.22) / 1000000)+350) / 10,

[0293] Where "*" represents multiplication (x).

[0294] These five radiomic features were then used to construct a radiomic signature of adipose tissue health, FatHealth (Figure 5(a)). FatHealth was independent of body mass index (r=0.068, p=0.307) and was positively correlated with circulating levels of C-reactive protein (CRP, r=0.315, p=0.0006). In the derivation cohort - with available adipose tissue biopsies - FatHealth was significantly associated with the presence of risk factors (smoking, arterial hypertension, and diabetes), reflecting their cumulative deleterious effects on adipose tissue biology (Figure 5(c)). This association was also significant in the test cohort, suggesting that FatHealth can be used as a non-invasive measure of adipose tissue health (Figure 5(d)).

[0295] Validation of a radiomics score for adipose tissue inflammation using positron emission tomography

[0296] In an independent cohort of 40 patients undergoing PET / CT imaging, subcutaneous adipose tissue was measured. 18 We investigated the association between F-FDG uptake and FatHealth calculated by radiomic feature extraction of subepithelial fat on CT scans. FatHealth was strongly positively correlated with adipose tissue FDG uptake as assessed by target-to-background ratio (TBR, Figure 5(e)) and had excellent diagnostic accuracy for high adipose tissue inflammation (i.e., TBR in the highest tertile, Figure 5(f)). 18 Validation of F-FDG PET / CT imaging, the gold standard method for assessing tissue inflammation in vivo, confirms its value as a noninvasive biomarker of adipose tissue inflammation.

[0297] Validation of the Alternative Radiomics Signatures of the Invention

[0298] The above discussion indicates that the radiomic signature calculated based on the five radiomic features listed in Table 2 (i.e., FatHealth) is an effective measure of adipose tissue dysfunction. In order to verify the performance of the alternative radiomic signatures of the present invention comprising different selections of radiomic features, a series of several different radiomic features were tested for their ability to identify adipose tissue dysfunction (previously defined as a state of high inflammation and low differentiation, i.e., high TNFA expression combined with low FABP4 expression). The results are shown in Tables 4 and 5 below, where Nagelkerke's pseudo R 2 Measures that discriminate between models of adipose tissue dysfunction are provided.

[0299] In Table 4 and Table 5, the model performance of the group A cohort is presented. In each of Examples 1-3, the current prior art model (including age, sex, hypertension, dyslipidemia, smoking, diabetes, body mass index, HOMA-IR index and subcutaneous fat mass) is gradually supplemented by radiomics signatures, and the radiomics features include gradually more radiomics features from different groups of Table 3. Therefore, in addition to those radiomics signatures listed in the previous row, each progressive step of each example corresponds to including the indicated radiomics features into the radiomics signature. The model of Example 1 only includes the "original" radiomics features listed in Table 2. In Example 2, each original radiomics feature in the original radiomics features has been replaced by the radiomics feature with the largest collinearity therewith, and in Example 3, each original radiomics feature in the original radiomics features has been replaced by the radiomics feature with the smallest collinearity therewith (see Table 3). As can be seen in Table 4, replacing the collinear equivalents with the original radiomic signature still yields a radiomic signature that provides incremental value over the current state of the art in differentiating adipose tissue dysfunction.

[0300] In Example 4, each of the original radiomic signatures in the "original" radiomic signatures of Table 2 has been replaced by an alternative radiomic signature from the same cluster, as identified using a hierarchical clustering algorithm (see Table 1). Again, Example 4 demonstrates that replacing each of the original radiomic signatures with an alternative radiomic signature in the same cluster still produces a radiomic signature that provides incremental value over the current state of the art in distinguishing adipose tissue dysfunction.

[0301] Table 4: Performance of alternative labels constructed from collinear equivalents

[0302]

[0303]

[0304] Table 5: Performance of alternative labels constructed from different features of the same cluster

[0305]

[0306] Thus, the data presented in Tables 4 and 5 demonstrate that, regardless of which features are selected from each of the identified clusters or groups, the radiomic signatures of the present invention provide improved discrimination of adipose tissue dysfunction over the current state of the art.

[0307] The radiomic signatures of the present invention were also derived for other fat depots, specifically breast and visceral adipose tissue, which are also useful targets for identifying adipose tissue dysfunction associated with metabolic health. The optimal radiomic signatures for breast and visceral adipose tissue are as follows:

[0308] Tags

[0309] (((-1264629630*large dependency low gray level emphasis)

[0310] +(259259.2593*average_LLL)

[0311] +(101629629.6*dependency entropy)

[0312] +(-121148148.1*Run Entropy)

[0313] +(TotalEnergy_HHL)) / 100000000)

[0314] +50,

[0315] Wherein "*" represents multiplication (x). Therefore, the radiomic signature of the present invention can also be applied to other fat depots, such as breast and visceral adipose tissue.

[0316] Results Overview

[0317] The radiomic signatures of the present invention can non-invasively identify adipose tissue dysfunction itself and characteristics of adipose tissue dysfunction, such as fibrosis and inflammation. Therefore, the radiomic signatures of the present invention can be used to predict the risk of developing metabolic disorders associated with adipose tissue dysfunction, such as diabetes.

[0318] Surprisingly, the radiomic signature need not consist of the radiomic features that are most strongly independently associated with adipose tissue dysfunction. Rather, it is actually advantageous to include radiomic features selected from different "clusters" of related or similar radiomic features, rather than only those radiomic features that are individually most associated with AT dysfunction. Furthermore, the "raw" radiomic features can be replaced with collinear equivalents while still providing an effective signature indicative of adipose tissue dysfunction.

[0319] A particularly attractive aspect of the present invention is that the present invention can be performed on historical medical imaging data that has been previously collected. The signatures of the present invention can be derived and calculated based on the historical imaging data, so the present invention provides a convenient tool for evaluating a large number of patients without having to perform further scans. Therefore, the method of the present invention does not have to include the step of collecting medical imaging data, and can be performed based on post-hoc analysis of existing medical imaging data.

Claims

1. A method for characterizing a region of interest comprising adipose tissue, the method comprising calculating a value of a radiomics signature of the region of interest using medical imaging data; wherein the radiomic signature is calculated based on measured values ​​of a plurality of radiomic features of the region of interest, the measured values ​​of the radiomic features being calculated based on the medical imaging data; in, The medical imaging data is computed tomography data; and Wherein the radiomic signature provides a measure of the texture of adipose tissue in the region of interest. 2 . The method of claim 1 , wherein at least one of the radiomic features provides a measure of texture of the region of interest.

3. The method of claim 1, wherein the radiomic signature is indicative of adipose tissue dysfunction. The method of claim 1 , wherein the radiomic signature predicts the likelihood that the subject will suffer from a metabolic disorder.

5. The method of claim 1, wherein the plurality of radiomic features comprises at least two radiomic features selected from the radiomic features of Group 1 to Group 5, wherein the at least two radiomic features are each selected from a different group, and wherein: Group 1 consisted of: large dependence low gray level emphasis, sum mean, joint mean, high gray level emphasis, short run high gray level emphasis, autocorrelation, and high gray level run emphasis; Group 2 consists of: total energy HHL, energy HHL, size zone non-uniformity HHH, size zone non-uniformity HLH, size zone non-uniformity LHH, normalized gray level non-uniformity HHL_GLDM, inverse variance HHL, normalized gray level non-uniformity HHL_GLSZM, size zone non-uniformity LHL, run length non-uniformity HHH, size zone non-uniformity HLL, dependence non-uniformity HHL, normalized gray level non-uniformity HHH_GLDM, inverse moment HHH, inverse difference HHH, non-uniformity HHH, related information measure 2, inverse variance HHH, total energy LHL, size zone non-uniformity, sphericity, size zone non-uniformity LLH, normalized run length non-uniformity HH L, normalized gray level non-uniformity HHH_GLSZM, normalized dependency non-uniformity HHL, zone percentage HHL, small dependency focus HHL, normalized size zone non-uniformity HHL, small area focus HHL, correlation, dependency non-uniformity LHH, total energy LHH, short run focus HHH, normalized run length non-uniformity HHH, dependency non-uniformity LHL, inverse difference LHH, inverse moment LHH, small dependency focus HHH, dependency variance HLH, run percentage HHH, zone percentage HHH, inverse difference LHL, dependency non-uniformity HLH, inverse variance LHH, large dependency focus HLH, normalized run length non-uniformity HLH, short run focus HLH, normalized Normalized dependency non-uniformity HLH, difference variance, long run focus HLH, energy LHL, contrast_GLCM, run percentage HLH, joint entropy LHH, difference entropy, small dependency focus HLH, normalized size area non-uniformity HLH, small area focus HLH, difference entropy LHH, difference mean, run variance HLH, inverse difference moment, inverse difference, inverse variance, run entropy LHH, entropy LHH, area percentage HLH, sum entropy LHH, normalized inverse difference moment, normalized inverse difference, small dependency low gray level focus, energy LHH, joint entropy LHL, small area low gray level focus, intensity, dependency non-uniformity LLH, run length non-uniformity HHL, normalized gray level non-uniformity LHL_GLDM, busy Looseness, normalized dependency non-uniformity LHH, small dependency focus LHH, coarseness, area percentage LHH, normalized dependency non-uniformity LHL, run length non-uniformity HLH, small dependency focus LHL, joint entropy, small area focus LHL, normalized size area non-uniformity LHL, short run focus HLL, run length non-uniformity LHH, normalized run length non-uniformity HLL, large dependency focus HLL, long run focus HLL, run percentage HLL, dependency variance HLL, dependency non-uniformity HLL, area percentage HLL, dependency non-uniformity, dependency non-uniformity HHH, normalized dependency non-uniformity LLH, low gray level area focus and short run low gray level focus; Group 3 consists of: Dependency entropy, Zone entropy, Correlation, Correlation information measure 2, Zone percentage HHH, Small dependency focus HHH, Run percentage HHH, Normalized dependency non-uniformity LHL, Small dependency focus LHL, Zone percentage HHL, Small dependency high gray level focus, Small dependency focus HHL, Normalized dependency non-uniformity HHL, Short run focus HHH, Normalized run length non-uniformity HHH, Normalized run length non-uniformity HHL, Normalized size zone non-uniformity LHL, Small area focus LHL , Normalized size area non-uniformity HHL, Small area focus HHL, Area percentage HLH, Run variance HLH, Small dependence focus HLH, Long run focus HLH, Normalized size area non-uniformity LLH, Small area focus LLH, Normalized size area non-uniformity HLH, Small area focus HLH, Run percentage HLH, Normalized run length non-uniformity HLH, Short run focus HLH, Large dependence focus HLH, Area percentage LHH, Small dependence focus LLH, Small dependence focus LHH, Normalized dependence non-uniformity Uniformity HLH, Zone Percentage LLH, Dependence Variance HLH, Normalized Dependence Non-Uniformity LLH, Normalized Dependence Non-Uniformity LHH, Inverse Difference LHL, Inverse Moment HHH, Inverse Difference HHH, Uniformity HHH, Normalized Gray Level Non-Uniformity LHL_GLSZM, Normalized Gray Level Non-Uniformity HHH_GLDM, Normalized Gray Level Non-Uniformity HHL_GLDM, Joint Entropy LHL, Normalized Inverse Difference, Normalized Inverse Moment, Long Run Focus HLL, Zone Percentage HLL, Inverse Difference, Inverse Moment , Difference Mean, Run Percentage HLL, Large Dependence Focus HLL, Contrast_GLCM, Short Run Focus HLL, Normalized Run Length Non-uniformity HLL, Inverse Variance, Inverse Variance HHL, Difference Entropy, Dependence Variance HLL, Inverse Difference Moment LHH, Difference Variance, Entropy LHH, Inverse Variance LHH, Inverse Difference LHH, Inverse Variance HHH, Sum Entropy LHH, Joint Entropy LHH, Normalized Gray Level Non-uniformity HHH_GLSZM, Normalized Gray Level Non-uniformity HHL_GLSZM, Difference Entropy LHH, and Run Entropy LHH; Group 4 consists of mean LLL, median LLL, and energy LLL; and Group 5 consists of run entropy. 6 . The method of claim 5 , wherein the at least two radiomic features include at least two of the following: large dependency low gray level emphasis, total energy HHL, dependency entropy, mean LLL, and run entropy.

7. The method of claim 5, wherein the at least two radiomic features consist of: Large dependence low gray level emphasis, total energy HHL, dependence entropy, mean LLL and run length entropy.

8. The method of claim 5, wherein the at least two radiomic features are selected from the radiomic features of clusters A to C, wherein the at least two radiomic features are each selected from a different cluster, and wherein: Cluster A consisted of the radiomic features of groups 1 to 3; Cluster B consists of the radiomics features of Group 4; and Cluster C consisted of the radiomic features of group 5.

9. The method of claim 1, wherein the plurality of radiomic features comprises at least two radiomic features selected from the radiomic features of clusters A to C, wherein the at least two radiomic features are each selected from a different cluster, and wherein: Cluster A is composed of: large dependence low gray level emphasis, run entropy, dependence entropy, cluster shading, skewness, run variance HLH, number of voxels, and gray level non-uniformity; Cluster B consists of: mean LLL, median LLL, joint mean, median, complexity, long run high gray level focus, sphericity, kurtosis, and roughness HHH; and Cluster C consists of the following: major axis, small dependent high gray level focus, minor axis, energy LLL, maximum 2D diameter row, long run low gray level focus, total energy HHL, normalized dependent non-uniformity HHH, contrast_GLCM, surface volume ratio, and entropy, normalized size area non-uniformity HHH, cluster prominence LHH, contrast LLH_GLCM and energy LHL.

10. The method of claim 5, wherein the at least two radiomic features include at least five radiomic features.

11. The method of any one of claims 1 to 10, further comprising predicting the subject's risk of developing a metabolic disorder based on at least the value of the radiomic signature.

12. The method of claim 11, wherein the metabolic disorder is diabetes or insulin resistance.

13. The method of any one of claims 1 to 10, further comprising determining whether the subject suffers from adipose tissue dysfunction based on at least the value of the radiomic signature.

14. A method for deriving a radiomic signature indicative of adipose tissue dysfunction, the method comprising using a radiomics dataset to construct a radiomic signature indicative of adipose tissue dysfunction, the radiomic signature being calculated based on a second plurality of radiomic features; wherein the dataset comprises values ​​of a first plurality of radiomic features obtained from medical imaging data of a region of interest, the region of interest comprising adipose tissue of each of a plurality of individuals, the plurality of individuals comprising a first group of individuals identified as having adipose tissue dysfunction and a second group of individuals identified as not having adipose tissue dysfunction; in, The medical imaging data is computed tomography data; wherein the second plurality of radiomic features are selected from the first plurality of radiomic features; and The radiomic signature is constructed to provide a measure of the texture of adipose tissue in the region of interest.

15. The method of claim 14, wherein each of the individuals is identified as having or not having adipose tissue dysfunction based at least in part on markers of adipose tissue dysfunction.

16. The method of claim 15, wherein the radiomics dataset further comprises the marker of adipose tissue dysfunction.

17. The method of claim 15, wherein the marker comprises the presence or absence of a metabolic disorder, or wherein the marker comprises a phenotypic characteristic of adipose tissue associated with adipose tissue dysfunction, or wherein the marker comprises a surrogate marker for a phenotypic characteristic of adipose tissue associated with adipose tissue dysfunction.

18. The method of any one of claims 14 to 17, further comprising identifying a first subset of the first plurality of radiomic features that are not collinear with each other, the second plurality of radiomic features comprising at least two radiomic features, each of the at least two radiomic features being selected as belonging to or being collinear with a different radiomic feature of the first subset.

19. The method of claim 18, further comprising identifying a second subset of radiomic features from the first subset using a feature selection algorithm, wherein the at least two radiomic features are selected to be different radiomic features belonging to or collinear with the second subset.

20. The method of claim 18, wherein each of the at least two radiomic signatures is selected to be stable.

21. The method of claim 20, wherein all unstable radiomic features are removed from the first plurality of radiomic features prior to selecting the second plurality of radiomic features.

22. The method of any one of claims 14 to 17 and 19 to 21, wherein the radiomic signature is constructed to be associated with adipose tissue dysfunction.

23. The method of any one of claims 14 to 17 and 19 to 21, wherein constructing the radiomics signature is performed using a machine learning algorithm.

24. The method of any one of claims 14 to 17 and 19 to 21, wherein at least one of the second plurality of radiomic features provides a measure of texture of the region of interest.

25. The method of any one of claims 14 to 17 and 19 to 21, further comprising a system configured to calculate the value of the radiomic signature of a patient.

26. The method of any one of claims 14-17 and 19-21, further comprising characterizing a region of interest of the patient by calculating the value of the radiomic signature of the region of interest of the patient, wherein the region of interest of the patient comprises adipose tissue.

27. The method of any one of claims 1 to 10, 14 to 17, and 19 to 21, wherein the adipose tissue is subcutaneous adipose tissue, visceral adipose tissue, or breast adipose tissue.

28. The method of any one of claims 1 to 10, 14 to 17, and 19 to 21, wherein the adipose tissue comprises non-cardiovascular adipose tissue.

29. A system configured to perform the method according to any one of claims 1 to 28.

Images