Radiomics and deep learning features of pericoronary adipose tissue on non-contrast computerized tomography for predicting non-calcified plaques

Abstract

Background

Inflammation of coronary arterial plaque is considered a key factor in the development of coronary heart disease. Early the plaque detection and timely treatment of the atherosclerosis could effectively reduce the risk of cardiovascular events. However, there is no study combining radiomics with deep learning techniques to predict non-calcified plaques (NCP) in coronary artery at present.

Objective

To investigate the value of combination of radiomics and deep learning features based on non-contrast computerized tomography (CT) scans of pericoronary adipose tissue (PCAT), integrating with clinical risk factors of patients, in identifying coronary inflammation and predicting the presence of NCP.

Methods

The clinical and imaging data of 353 patients were analyzed. The region of interest (ROI) of PCAT was manually outlined on non-contrast CT scan images, like coronary CT calcium score sequential images, then the radiomics and deep learning features in ROIs were extracted respectively. In training set (Center 1), after performing feature selection, radiomics and deep learning feature models were established, meanwhile, clinical models were built. Finally, combined models were developed out via integrating clinical, radiomics, and deep learning features. The predictive performance of the four feature model groups (clinical, radiomics, deep learning, and three combination) was assessed by seven different machine learning models through generation of receiver operating characteristic curves (ROC) and the calculation of area under the curve (AUC), sensitivity, specificity, and accuracy. Furthermore, the predictive performance of each model was validated in an external validation set (Center 2).

Results

For the single model comparation, eXtreme Gradient Boosting (XGBoost) showed the best performance among the clinical model group in the validation set. And Random Forest (RF) exhibited the best indicative performance not only among the radiomics feature group but also in the deep learning feature model group. What's more, among the combined model group, RF still displayed the best predictive performance, with the value of AUC, sensitivity, specificity, and accuracy in the validation set are 0.963, 0.857, 0.929, and 0.905.

Conclusion

The RF model in the combined model group based on non-contrast CT scan PCAT can predict the presence of NCP more accurately and has the potential for preliminary screening of the NCP.

Keywords

pericoronary adipose tissue computed tomography machine learning radiomics deep learning

Introduction

Coronary artery disease (CAD), a prevalent globally cardiovascular disorder, usually leads to myocardial ischemia and sudden death and threatens the human life, health, and happiness.¹ CAD is an inflammatory arterial disease and closely linked to lipid accumulation and metabolic alteration. Triggered by various risk factors, atherosclerosis serves as the primary pathological manifestation of CAD. Beyond dyslipidemia, inflammation also plays a vital role in the pathophysiological mechanisms of atherosclerosis and its associated ischemic events. PCAT is defined as the adipose tissue locating in a distance from the vessel exterior wall equivalent to the respective vessel diameter and exerting its function via paracrine secretion. Simultaneously, atherosclerosis of the vessel wall can inhibit the maturation of adipocytes and promote the accumulation of lipids in the PCAT,² which creates gradient discrepancy between the lipid-rich and less aqueous phase belonged to a non-diseased artery and lipid-poor and more aqueous phase belonged to an inflamed artery (Figure S1), and further causes the PCAT attenuation.^3–6 Serving as a sensitive and dynamic biomarker for coronary artery inflammation, the fat attenuation index (FAI) is capable of identifying PCAT attenuation. It is shown that FAI values in the proximal right coronary artery (RCA) can reflect pathological and physiological changes throughout the coronary arteries.⁷

So, timely discovery of coronary artery inflammation in PCAT will assist in the early prevention or administration of CAD. However, it is absent of specificity for common method to identify the coronary artery inflammation in PCAT (e.g., serum biomarkers). Additionally, for some special examination, the exorbitant expenses of PET scans, the low resolution of ultrasound imaging, and the inapplicability of magnetic resonance imaging (MRI) for special patients who with claustrophobia or metal implants all limit their application in clinical diagnosis and treatment.⁸ Coronary CT angiography (CCTA) examination has several disadvantages yet, including a higher risk of radiation exposure, potential allergic reactions caused by iodine-induced contrast agents, adverse effects on patients with renal insufficiency, long examination times, and high costs. The intravenous injection of contrast agents also requires needles, which may cause additional vascular damage and discomfort of patients. Thereby, it is restricted in clinical practice at a certain degree.⁹

Furthermore, the present studies at have discovered that iodine contrast agents can increase the attenuation of PCAT in inflammatory conditions, which affects the capability of risk prediction of PCAT. In contrast, non-contrast coronary CT calcium score sequences provide another safe and reliable imaging option. It avoids potential interference introduced by iodine contrast agents under inflammatory conditions and accurately reflects the real FAI. This is particularly valuable for assessing the true FAI of PCAT.¹⁰

Detection of NCP in the coronary arteries is crucial for cardiovascular risk assessment, because of they are prone to rupture and thrombosis formation.¹¹ In clinical practice, the diagnosis of CAD is usually a complex process which mainly dependents on clinicians’ experience and their subjective judgment of imageology.^12–14 For subtle symptoms, inexperienced young physicians may be at risk of misdiagnosis. Today, the integration of radiomics and deep learning technologies has brought innovation to precision medicine.^15,16 It can capture subtle pathological signs, such as fibrosis and microvascular remodeling that are difficult to be recognized by the traditional methods. It provides more comprehensive and objective diagnostic evidence through the analysis of machine learning models.

Therefore, in this study, radiomics and deep learning features of PCAT were extracted from coronary CT calcium score (CCS) sequence images. Combined with clinical features, we utilized seven machine learning models to establish 28 predictive models and divide them into four feature groups: clinical feature (Group 1), radiomics feature (Group 2), deep learning feature (Group 3), and combined clinical-radiomics-deep learning feature (Group 4). External validation data was used to validate those models. Through comprehensively comparing the forecast performance of different models with each group, we identified the most suitable model for predicting NCP. This model can be used for initially screening and indicating the further examination and treatment of suspicious patients, and significantly improving the efficiency of clinical resources and reducing healthcare expenses.^17,18

Methods

Study cohort

The research adhered to the principles of the Declaration of Helsinki (2013 revision) and received approval from the Institutional Review Board (IRB) of Qiqihar Medical University; therefore, consent for this retrospective analysis from individuals was not required. The training dataset was obtained from the Third Affiliated Hospital of Qiqihar Medical University (Center 1), while the external validation dataset was obtained from the Second Affiliated Hospital of Qiqihar Medical University (Center 2).

The clinical and imaging data of inpatients diagnosed with or suspected of coronary heart disease was collected from two affiliated hospitals of Qiqihar Medical University between January 2019 and June 2023. In the Radiology Picture Archiving and Communication System (PACS) of hospitals, patients who underwent CCTA with 'non-calcified plaques’ were categorized into the non-calcified plaques group and with 'no significant abnormalities in coronary arteries’ were divided into the control group. The inclusion criteria were as follow: ① Complete baseline data, including patient age, gender, diabetes, hypertension, smoking history, glucose levels, total cholesterol, triglycerides, total protein, albumin, high density lipoprotein cholesterol, low density lipoprotein cholesterol, apolipoprotein A-1, and apolipoprotein B-1; ② No history of iodine allergy; ③ No severe heart, lung, or kidney dysfunction. The exclusion criteria were as below: ① Patients who underwent coronary artery bypass surgery of the RCA; ② Patients with stents, cardiac pacemakers, metallic valve prostheses, or myocardial bridges in the RCA; ③ Presence of respiratory or motion artifacts in the CCS sequence images of the RCA. A total of 353 cases were finally involved, including the 249 cases in final training set (Center 1) and 104 cases in the external validation set (Center 2, Figure S2).

Ct image acquisition

Center 1 utilized the Apex GE Revolution 256 slice spiral CT for image acquisition, while Center 2 utilized the GE Revolution 256 slice spiral CT. For patients with a rapid heart rate, it was recommended to administer oral β- blockers to lower the heart rate to below 65 beats per minute. Patients were positioned supine with head-first entry into the scanner with arms raised and stretching. The scan range extended from the tracheal bifurcation to 1–2 cm below the diaphragm.

CT scanning parameters: the detector width was 38 mm, the rotating tube was 0.28 s, the pitch ranged from 0.20–0.35, the collimator width was 128 × 0.5 mm, the tube voltage was set to 120 kV, utilizing automatic modulation of tube current ranging between 300 to 600 mA. Following the plain scan, a dynamic contrast-enhanced CT scan was performed to minimize respiratory and motion artifacts. Respiratory training was conducted during the scan to improve image quality. Iopromide (370 mg/ml iodine concentration, Bayer Healthcare Limited) was administered via the elbow vein at 5 ml/s with an initial injection volume of 60–80 ml followed by 40–50 ml of saline solution. Retrospective electrocardiogram gating was utilized for cardiac gating. The slice thickness for CCS and CCTA was set 0.625 mm and reconstruction slice thickness was 1.25 mm. Diastolic and systolic images were reconstructed at 30–65% and 70–80% of the RR intervals respectively, and the optimal data was selected.

PCAT segmentation, radiomics and deep learning features extraction

Quantification of PCAT proximal to the RCA is a standardized and highly repeatable method and an useful imaging biomarker to assess the overall state of coronary inflammation in prior studies.^3,7,19,20 Therefore, we segmented the PCAT of RCA and extracted radiomics features and deep learning features. The original imaging data were loaded into the 3D Slicer software (version 5.2.2), where the manual segmentation of the ROI surrounding the right coronary artery was conducted by a radiologist possessing a decade of diagnostic experience. For the uncertainty, a senior radiologist will be in. Detailed methodology: firstly, using the 3D Slicer open-source software, PCAT in the range of 10–50 mm from the proximal opening of the RCA were segmented layer by layer on a CCS image with a layer thickness of 1.25 mm, a window width of 350 HU, and a window position of 50 HU. Secondly, using the “hollow” function to expand the PCAT to an equal distance as the RCA diameter and automatically erasing the RCA, and then manually removing non-adipose tissue in the ROI. At last, a three-dimensional volume of interest (VOI) for pericardial adipose tissue was obtained.

Before features extraction, the CT images underwent several preprocessing steps. Firstly, voxel resampling was performed to achieve isotropic voxels with dimensions of 0.5 mm × 0.5 mm × 0.5 mm to enable standardized analysis of CT images. Secondly, gray-level discretization was applied to convert the continuous image to discrete integer values with a bin width of 25. Finally, Laplacian of Gaussian (LoG) and wavelet filters were employed to extract high-frequency or low-frequency features.

Pyradiomics package was employed for radiomics feature extraction from the delineated VOI. Those features encompassed first-order statistics, shape-based features, and texture features including the gray-level dependence matrix (GLDM), gray-level co-occurrence matrix (GLCM), gray-level size zone matrix (GLSZM), gray-level run length matrix (GLRLM), and the neighboring gray tone difference matrix (NGTDM), as well as wavelet-based features.²¹

3D ResNet architecture was applied as the baseline model for extracting deep learning features. This architecture modifies the classic ResNet²² by replacing all two-dimensional convolutional layers with three-dimensional convolutional layers to process three-dimensional CT scan image data to capture the three-dimensional spatial features in the VOI. During the feature extraction process, global pooling is applied after the final convolutional layer to aggregate the features, which are subsequently input into machine learning models.

To evaluate the method stability and repeatability, 50 patient image sets were selected randomly from the training dataset two weeks later and were segmented independently for PCAT by another experienced CT diagnostician with 15 years of experience without knowledge of the initial results. Intraclass correlation coefficient (ICC) assessed the consistency between the segmentations by the two physicians, ensuring rigor and reliability of the entire analysis process. The overall workflow is shown in Figure 1.

Figure 1.

Overall workflow diagram.

Feature selection

Initially, we retained those radiomics and deep learning features having an ICC exceeding 0.80. Subsequently, Levene's test for equality of variances and a normality test were conducted on features of the training set comprised of 249 patients from the control group and the non-calcified plaques group to ascertain whether the assumptions of equal variance and normal distribution were met. If satisfied, the t-test was employed for comparison analysis, otherwise, the Welch's test was used. By comparing the two groups’ features, we selected those features with P < 0.05 for subsequent analysis steps. Secondly, we utilized Spearman rank correlation coefficients to calculate the correlation between features, one was retained if the absolute value of the correlation coefficient between two features was greater than 0.75. Finally, we identified the significance of features by analyzing the coefficients of predictive variables in the observed regression model using the least absolute shrinkage and selection operator (LASSO) analysis,²³ while determining the optimal penalty coefficient lambda through using 5-fold cross-validation method to obtain the final features selection result.

Development of radiomics, clinical, deep learning, and combined feature models

A univariate and multivariate logistic regression analysis was conducted on all clinical features in the training set, including age, gender, diabetes, hypertension, smoking, total cholesterol, triglycerides, total protein, high density lipoprotein cholesterol, low density lipoprotein cholesterol, apolipoprotein A-1, and apolipoprotein B-1. After screening, clinically features with significant differences were retained to construct the clinical feature models (Group 1). Simultaneously, after multi-step screening processes of radiomics and deep learning features were operated, suitable radiomics and deep learning features for predicting NCP were retained to establish the radiomics feature models (Group 2) and the deep learning feature models (Group 3). The final screened clinical features, radiomics features, and deep learning features were then combined to build combined feature models (Group 4). Seven models were utilized to authenticate the predictive performance of each group, including Logistic Regression (LR), Support Vector Machine (SVM), RF, Stochastic Gradient Descent (SGD), K-Nearest Neighbors (KNN), XGBoost, and Light Gradient Boosting Machine (LightGBM). In addition, ROC were used on the external validation set to calculate AUC and assess the accuracy of each model in predicting NCP.

Model interpretation and feature importance visualization

Machine learning models are often regarded as black boxes because of the difficulty in understanding the reasoning behind their predictions for specific patient cohorts. Therefore, this study introduced SHAP (Shapley Additive exPlanation) to elucidate the supposed mechanisms of high-performing machine learning models.²⁴ Proposed by Lundberg and Lee, SHAP provides a unified framework for interpreting machine learning predictions and offers a novel approach to explaining various black-box models. Its interpretability performance has been previously validated.^25,26 SHAP is capable for performance both local and global interpretability analyses and possesses a solid theoretical foundation compared to other methods.²⁷ In this study, SHAP was applied to interpret the final predictive model, aiming to identify the primary predictive factors influencing NCP in the patient cohort and to rank the importance of features accordingly.

Statistical analysis

Statistical analysis was performed with SPSS 26.0 and Python (version 3.9.0, https://www.python.org/) utilization. Patients were classified according to the presence or absence of NCP in RCA. For univariate analysis, qualitative data were compared between groups using the chi-square test and expressed as frequency (percentage). Quantitative data were first subjected to a normality test. If data conformed to or was close to normal distribution, an independent-samples t-test was used to compare differences between the two groups, data was described with mean ± standard deviation (SD). If the data did not conform to normal distribution, the non-parametric Mann-Whitney U test was employed, with results reported as median (interquartile range). Significant features identified in the univariate analysis were included in a multivariate logistic regression to ascertain independent risk factors, presented with odds ratios (OR) and 95% confidence intervals (CI). An external validation set comprising 104 patients from Center 2 was used to test the accuracy of the models. The discrimination ability of each model within the four groups was evaluated using ROC and AUC values. The classification performance of the best model was evaluated using a 2 × 2 confusion matrix. Decision curve analysis (DCA) visualized the clinical net benefit provided by the models.²⁸ The DeLong test method was employed to assess the significance of differences between the best model in the combined models and other models.²⁹ In this study, all statistical tests were two-tailed, with P < 0.05 considered statistically significant.

Results

The study population characteristics

The baseline data of 353 patients in the study were analyzed firstly. Significant differences were observed between the control and non-calcified plaques groups in terms of age, sex, diabetes, hypertension, smoking history, albumin level, total protein content, glucose level, high-density lipoprotein cholesterol, and apolipoprotein A-1 (P < 0.05). However, there was no significant differences among the groups in the assay of total cholesterol, triglyceride, and low density lipoprotein cholesterol (P > 0.05). Detailed baseline characteristics of the training set, external validation set, and all patients were showed in Table 1 and Table 2 respectively. Multivariate logistic regression analysis revealed that age, history of diabetes, smoking history, albumin and low-density lipoprotein cholesterol assay were independent risk factors for predicting the presence of atherosclerotic heart disease (P < 0.05, Table 3). Besides, albumin and high density lipoprotein cholesterol exhibited significant disparities between the training and validation groups (P < 0.05), while age, diabetes, and smoking history did not differ significantly between this two groups (P > 0.05, Table 4).

Table 1.

Baseline characteristics of patients in the training and validation sets.

	Training set					Validation set
Characteristic	Control group (N = 140)	Non-calcified plaques group (N = 109)	t/Z/χ²	P-value	Control group (N = 74)	Non-calcified plaques group (N = 30)	t/Z/χ²	P-value
Age	54.26 ± 10.58	60.85 ± 9.88	−5.023	0.000	56.50 ± 10.13	65.37 ± 8.93	−4.179	0.000
Gender			5.473	0.019			2.936	0.087
Male	60 (48.8%)	63 (51.2%)			26 (61.9%)	16 (38.1%)
Female	80 (63.5%)	46 (36.5%)			48 (77.45%)	14 (22.6%)
Diabetes	19 (21.3%)	70 (78.7%)	68.451	0.000	9 (32.1%)	19 (67.9%)	28.41	0.000
Hypertension	39 (33.3%)	78 (66.7%)	46.990	0.000	21 (51.2%)	20 (48.8%)	13.104	0.000
Smoking	17 (22.7%)	58 (77.3%)	49.108	0.000	10 (34.5%)	19 (65.5%)	26.347	0.000
Glucose	5.67 (5.36, 6.11)	6.21 (5.55, 7.52)	−4.037	0.000	5.88 (4.99, 6.428)	6.36 (5.61, 7.82)	−3.078	0.002
TC	5.15 (4.74, 5.63)	4.84 (4.2, 5.38)	−2.876	0.004	4.88 ± 1.05	5.17 ± 1.12	−1.278	0.204
Triglyceride	1.73 (1.23, 2.34)	1.72 (1.22, 2.51)	−0.558	0.577	1.87 (1.33, 2.56)	1.85 (1.06, 2.25)	−0.768	0.443
Total protein	74.30 ± 4.24	73.65 ± 4.60	1.157	0.248	67.23 (64.84, 71.26)	75.44 (68.39, 77.43)	−4.004	0.000
Albumin	46.31 (44.8, 47.89)	45.40 (44.3, 47.945)	−1.318	0.188	41.54 (39.68, 43.65)	45.92 (43.13, 47.73)	−4.144	0.000
LDL-C	3.25 (2.81, 3.66)	3.02 (2.56, 3.41)	−2.699	0.007	2.95 ± 0.77	3.30 ± 0.97	−1.915	0.058
HDL-C	1.29 (1.17, 1.43)	1.18 (1.08, 1.35)	−3.465	0.001	1.47 ± 0.36	1.26 ± 0.30	2.798	0.006
ApoA-1	1.32 ± 0.12	1.31 ± 0.13	0.763	0.446	1.09 (1.07, 1.20)	1.34 (1.09, 1.35)	−2.788	0.005
ApoB-1	0.99 (0.88, 1.08)	0.90 (0.82, 1.00)	−3.993	0.000	0.85 (0.44, 1.09)	0.84 (0.79, 1.16)	−1.579	0.114

TC, Total cholesterol; LDL-C, Low-Density Lipoprotein Cholesterol; HDL-C, High density lipoprotein cholesterol; ApoA-1, apolipoprotein A-1; ApoB-1, apolipoprotein B-1.

Table 2.

Comparison of baseline characteristics between Control and Non-calcified plaques group in all study subjects.

Characteristic	Control group (N = 214)	Non-calcified plaques group (N = 139)	P-value
Age	55.03 ± 10.46	61.83 ± 9.83	0.000
Gender			0.002
Male	87 (52.4%)	79 (47.6%)
Female	127 (67.9%)	60 (32.1%)
Diabetes	28 (13.1%)	89 (64.0%)	0.000
Hypertension	59 (27.6%)	98 (70.5%)	0.000
Smoking	27 (12.6%)	77 (55.4%)	0.000
Glucose	5.7 (5.26, 6.4)	6.21 (5.57, 7.57)	0.000
Total cholesterol	5.04 (4.58, 5.59)	4.96 (4.22, 5.55)	0.132
Triglyceride	1.78 (1.24, 2.39)	1.75 (1.21, 2.36)	0.966
Total protein	72.18 (67.94, 75.73)	73.62 (70.48, 76.82)	0.010
Albumin	45.30 (42.09, 47.15)	45.50 (44.20, 47.80)	0.021
LDL-C	3.16 (2.70, 3.66)	3.04 (2.58, 3.62)	0.244
HDL-C	1.31 (1.19, 1.54)	1.19 (1.08, 1.36)	0.000
ApoA-1	1.25 (1.14, 1.37)	1.31 (1.19, 1.41)	0.022
ApoB-1	0.92 (0.84, 1.08)	0.89 (0.82, 1.03)	0.148

LDL-C, Low-Density Lipoprotein Cholesterol; HDL-C, High density lipoprotein cholesterol; ApoA-1, apolipoprotein A-1; ApoB-1, apolipoprotein B-1.

Table 3.

Multivariable logistic analysis of independent risk factors for Non-calcified plaques.

Characteristic	B	P	OR	95% CI
Age	0.08	0.000	1.084	1.048–1.121
Gender	0.555	0.077	1.742	0.942–3.222
Diabetes	1.828	0.000	6.222	3.114–12.431
Hypertension	0.231	0.502	1.26	0.642–2.471
Smoking	1.652	0.000	5.217	2.699–10.086
Total protein	0.023	0.481	1.023	0.96–1.09
Albumin	0.11	0.046	1.116	1.002–1.243
ApoA-1	0.596	0.605	1.815	0.19–17.316
HDL-C	−1.369	0.008	0.254	0.093–0.698

ApoA-1, apolipoprotein A-1; LDL-C, Low Density Lipoprotein Cholesterol.

Table 4.

Comparison of baseline features between training set and validation set.

Characteristic	Training set (N = 249)	Validation set (N = 104)	P-value
Age	57.14 ± 10.77	59.06 ± 10.56	0.127
Diabetes	89 (35.7%)	28 (26.9%)	0.109
Smoking	75 (30.1%)	29 (27.9%)	0.674
Albumin	46.10 (44.4, 47.90)	42.38 (40.28, 45.63)	0.000
HDL-C	1.26 (1.11, 1.42)	1.34 (1.17, 1.64)	0.001

HDL-C, High density lipoprotein cholesterol.

Clinical feature models

When univariate logistic regression analysis on the training set finished, we identified 10 clinical factors associated with predicting NCP. Subsequently, we determined three independent predictors related to NCP through multivariate logistic regression analysis, they are age, diabetes, and smoking history. Then, these factors were incorporated into clinical models and a detailed comparison of the model's predictive performance was conducted (Figure 2, Table S1). Among those clinical feature models, XGBoost exhibited the optimal predictive performance, achieving an AUC value of 0.940, sensitivity of 0.780, specificity of 0.935, and accuracy of 0.867 in the training set. For the validation set, the corresponding value of AUC, sensitivity, specificity, accuracy was 0.885, 0.810, 0.810, and 0.810 respectively.

Figure 2.

Results of clinical features in training sets and external validation sets by using SVM, RF, SGD, KNN, LightGBM, XGBoost, and Logistic Regression.

Radiomics feature models

1316 radiomics features were extracted from the CT images, but after features filtering, only 25 relevant radiomics features for predicting NCP were picked out (Figure 3A, B, Table S2). These features were incorporated into the seven machine learning models mentioned above, and detailed comparison of the predictive performance of those models were carried out (Figure 3C, D, Table S3). We found that among those group of radiomics feature models, the RF model exhibited the best predictive performance, achieving AUC values of 0.935, sensitivity of 0.788, specificity of 0.979, and accuracy of 0.851 for the training set. For the external validation set, it still showed that value of AUC, sensitivity, specificity, accuracy was separately 0.908, 0.756, 0.986, 0.863.

Figure 3.

Using Lasso for radiomics feature selection and the results of radiomics feature models in training sets and external validation sets. (A) Selection of penalty parameter lambda using 5-fold cross-validation. (B) The convergence plot of feature coefficients in LASSO regression, with the curve representing the change of feature coefficients with λ (0.003727). (C, D) Results of radiomics features in training sets and external validation sets via using SVM, RF, SGD, KNN, LightGBM, XGBoost, and Logistic Regression.

Deep learning feature models

There were 2048 deep learning features extracted from the images. After features screening, 20 deep learning features were supposed to involve in the prediction of NCP (Fiure. 4A, B, Table S4). These features were incorporated into seven machine learning models, and the predictive performance of these models were compared in detail at the same (Figure 4C, D, Table S5). We found that the RF model achieved the best predictive presentation among the deep learning feature models, achieving an AUC value of 0.979, sensitivity of 0.862, specificity of 0.964, and accuracy of 0.919 for the training set. On the validation set, this model still attained an AUC value of 0.910, sensitivity of 0.762, specificity of 0.952, and accuracy of 0.889.

Figure 4.

Using Lasso for deep learning feature selection and results of deep learning feature models in training sets and external validation sets. (A) Selection of penalty parameter lambda using 5-fold cross-validation. (B) The convergence plot of feature coefficients in LASSO regression, with the curve representing the change of feature coefficients with λ (0.006551). (C, D) Results of deep learning features in training sets and external validation sets by using SVM, RF, SGD, KNN, LightGBM, XGBoost, and Logistic Regression.

Combined clinical-radiomics-deep learning feature models

To explore the higher efficiency and more accurate predictive model, we integrated clinical, radiomics, and deep learning features and constructed combined feature models. After the combination, the predictive performance of these combined feature models was meticulously compared each other (Figure 5, Table S6). The results indicated that among these combined features models group, RF presented the best predictive performance, with an AUC value of 0.998, sensitivity of 0.945, specificity of 0.998, and accuracy of 0.976 for the training set. For the validation set, the model also achieved same excellent performance for prediction with an AUC value of 0.963, sensitivity of 0.857, specificity of 0.929, and accuracy of 0.905.

Figure 5.

Results of combined features on training sets and external validation sets through using SVM, RF, SGD, KNN, LightGBM, XGBoost, and Logistic Regression.

Comparison of models and visualization of feature importance

Depended on those above results (Table 5), the RF model in the combined feature group achieved an AUC value of 0.963, significantly outperforming the predictive performance of the best model in the clinical feature group (AUC = 0.885), the best model in the radiomics feature group (AUC = 0.908), and the best model in the deep learning feature group (AUC = 0.910). It illustrated that the RF model is the best predictive model in combined models. The DeLong test results (Table 6) demonstrated that on both the training set and the external validation set, the P-values between the best model in combined models and in each single models (clinical model, radiomics model, deep learning model) were all statistically significantly different. It illustrates that the confusion matrix of the RF model using the combined features set on both the training set and the external validation set, providing a more intuitive view of the classification effect (Figure S3). The diagonal elements represent correct predictions and match with the actual labels, while the off-diagonal elements signify incorrect predictions. Darker shades indicate better classification performance and higher accuracy. To explain the effect of the selected variables on the model's prediction of NCP in greater depth, the top 20 risk factors with the greatest impact on the prediction of the Random Forest model can be clearly seen through the average absolute SHAP values shown in Figure S4, which provides a deeper explanation of the impact of selected variables on the model's prediction of non-calcified plaques. Besides, our study plotted decision curves of the RF model with the highest predictive performance across the combined feature group, and the results showed that contrast to other models, the highest clinical benefit was reaped by using combined feature group models (Figure S5).

Table 5.

Summary of optimal models for each group.

Group	Optimal Model	Category	AUC	Sensitivity	Specificity	Accuracy
Clinical feature	XGBoost	Training set	0.940	0.780	0.935	0.867
Clinical feature	XGBoost	Validation set	0.885	0.810	0.810	0.810
Radiomics feature	Random Forest	Training set	0.935	0.788	0.979	0.851
Radiomics feature	Random Forest	Validation set	0.908	0.756	0.986	0.863
Deep learning feature	Random Forest	Training set	0.979	0.862	0.964	0.919
Deep learning feature	Random Forest	Validation set	0.910	0.762	0.952	0.889
Combined feature	Random Forest	Training set	0.998	0.945	0.998	0.976
Combined feature	Random Forest	Validation set	0.963	0.857	0.929	0.905

XGBoost, eXtreme Gradient Boosting.

Table 6.

Delong test results comparing combined group's best model with each sub-group's best model.

	Training set	Validation set
Clinical feature model	P < 0.05	P < 0.05
Radiomics feature model	P < 0.05	P < 0.05
Deep learning feature model	P < 0.05	P < 0.05

Discussion

CAD has been a major challenge in public health due to a continuously increasing trend in morbidity and mortality.³⁰ So, it is urgent to develop a screening tool to predict NCP both efficiently and simply and aim to achieve early diagnosis and treatment of the disease, improving patient prognosis. Therefore, this study adopted an innovative approach to extract radiomics and deep learning features by combining PCAT of CT plain images through ultrathin voxel technology and integrated them with clinical data to establish four feature model groups, every feature model group contained seven kinds of models. The study totally evaluated the efficacy of 28 different modeling schemes in predicting NCP and used independent external datasets to validate the models for searching the optimal prediction models and ensuring accuracy. We found that in the combined model group, the RF model exhibited excellent prediction accuracy and stability, which not only strongly proved potential of the model application in the initial screening stage of NCP, but also effectively assist suspected cases to receive in-depth medical examination.

As many researching achievement and progresses have been gained in radiomic study, it is possible to predicting CAD by radiomics application. KOLOSSVARY et al.³¹ conducted an ex vivo study on seven donor hearts, scanning the coronary arteries. Those results showed that radiomics exhibited higher diagnostic efficacy (AUC = 0.73) compared to relying solely on CCTA visual assessment and histogram analysis for identifying advanced atherosclerotic lesions. However, compared to the predictive performance of the best model in this study, the performance of their model was unsatisfactory because of it based on a single radiomics feature technology. On the other hand, ZREIK et al.³² focused on effectively identifying patients with significant coronary artery stenosis, achieved automatic quantification and classification of the degree of coronary stenosis, thus further expanding the study of CAD. HOMAYOUNIEH et al.³³ also explored the relationship between coronary artery calcification and stenosis depended on radiomics. Despite significant advances are obtained, most studies are just based on CCTA technology and do not account for the attenuation effect on pericoronary fat after using contrast agents and its potential risks, which limits practical application in large-scale population screening.

Tao et al.³⁴ explored the diagnostic value of the radiomics model of CT plain scan PCAT for NCP and found that the radiomics model had higher predictive performance (AUC = 0.750) compared with the FAI value. However, compared to our study, the result was inferior, possibly due to using a single radiomics model without combining patient clinical characteristics, which probably results in poor model fit. Jiang et al.³⁵ found that gender, age, hypertension, diabetes, smoking, clinical diagnosis, and the average CT value of pericoronary fat were independent risk factors for predicting NCP. The combined model outperformed the single radiomics one by incorporating both clinical factors and radiomics features. Our study also showed that age, diabetes, and smoking history were independent risk factors for predicting NCP and fit with the previous study. Additionally, this study found that the model added with deep learning features performed better than a single radiomics model.

Previous studies have relied on radiomics features alone to predict the presence of NCP, however, this method is limited by the manually defined features, which can only reflect a relatively limited amount of image details, thus restricting the predictive performance of the machine learning model. Those latest advancements in deep learning are poised to make significant contributions to the field of medical imaging. By making models to learn from data autonomously, this deep learning eliminates the constraints of manual feature definitions. Specifically, the usage of transfer learning allows for the transfer of prior knowledge of image features and their application in medical imaging ultimately improving the generalization performance of the model.³⁶

Although lots of researches have reported the different models relying on single feature or two features combination, currently, there is no research combinate clinical, radiomics, and deep learning features to predict NCP. This study employed a 3D ResNet deep learning network for extracting deep features, selecting 20 features for further analysis. Additionally, we retained 3 features from clinical parameters and 25 features from radiomics, and integrating them together into one model for modeling. The results revealed that the RF model exhibited superior performance in predicting NCP in the combined clinical-radiomics-deep learning feature group. Specifically, the training set achieved an AUC value of 0.998, sensitivity of 0.945, specificity of 0.998, and accuracy of 0.976. Similarly, the validation set recorded an AUC value of 0.963, sensitivity of 0.857, specificity of 0.929, and accuracy of 0.905. These findings suggest a potential correlation between PCAT and the presence of NCP and aligning with the outside-in pathogenesis theory of coronary artery atherosclerosis. This theory posits that inflammatory adipocytes in PCAT promote atherosclerotic progression by releasing adipocytokines.³⁷ It is speculated that inflammation can lead to morphological changes in PCAT, such as increased water retention in adipocytes and alterations in vascular structure. These subtle changes can be captured through radiomics and deep learning features, which provides a novel perspective for a deeper understanding of the underlying biological mechanisms of coronary artery atherosclerosis. The RF model based on non-contrast enhanced PCAT scans, exhibited remarkable accuracy in identifying NCP and significantly outperformed models that was constructed solely with clinical features, radiomics features, or deep learning features.

Reasons for the improvement may include: firstly, the use of 0.5 mm ultra-thin voxel layers can provide more detailed and accurate image information. Secondly, this study adds LoG and wavelet features when extracting radiomics features, and combines deep learning features and clinical features to better fit the model. Thirdly, the RF model integrates the prediction results of multiple decision trees, reduces the overfitting risk, improves the generalization ability of the overall model, and has a good fitting ability for complex feature relationships, so that, it can better to get the correlation and information among them when combining radiomics, clinical and deep learning features.

Given the "black box" nature of advanced machine learning models, this study employed the SHAP algorithm for in-depth analysis. This algorithm can be applied to any type of machine learning model, enjoys the advantages of fast implementation for tree-based models, and guarantees consistency and local accuracy. By conducting interpretable analysis on the RF model with combined feature group, the top 20 factors influencing NCP were ranked and revealed that texture and geometry-based radiomics features are crucial for predicting the presence of NCP, which aligns with previous research findings.³⁸ These features can identify the complexity and heterogeneity of plaques interior components that cannot be fully captured by traditional CT measurements. Future studies should further explore the biological significance of these radiomics features to better understand the pathophysiology of NCP.³⁹ Previous research⁴⁰ has confirmed the potential mechanism of age's influence on NCP and suggests that age differences should be considered in clinical practice.

This study has some limitations. Firstly, it utilized a retrospective study design with a limited sample size potentially result in introducing selection bias. Our future research plans to increase the sample size to enhance results generalizability. Secondly, the study only explored the presence of NCP in the RCA, excluding the other two coronary arteries. Our future research will focus on investigation whether adipose tissue surrounding the RCA can predict the presence of NCP in each coronary artery and will establish a PCAT model for comparison with the other two coronary arteries.

The RF model in the clinical-radiomics-deep learning combined feature group predicted NCP significantly on PCAT images from non-contrast CT scans. On one hand, this approach eliminates the risk of allergic reactions that potentially caused by contrast agents. On the other hand, it facilitates to guide potential patients to make further examinations, consequently reducing economic burden and medical costs for patients and enhancing the positive detection rate.

In the future, it is expected that data from chest CT scans can be combined with artificial intelligence detection technology widely not just for predicting CAD but lung nodules. Using this method, artificial intelligence analysis could also be applied to PCAT, ultimately achieving the dual goals of coronary heart disease and lung cancer screening with once chest CT scan. This advancement will not only provide more convenient and efficient medical services to patients but also contribute to early disease detection, diagnosis, and treatment for making significant contributions to improving public health.

Conclusions

In summary, this study made a comprehensive evaluation of the RF model's ability to predict NCP on CT scans of PCAT in the clinical-radiomics-deep learning combined feature group. Those results demonstrate that this model has higher accuracy and more promising validating for predicting NCP in clinical applications.

Supplemental Material

sj-pdf-1-xst-10.1177_08953996241292476 - Supplemental material for Radiomics and deep learning features of pericoronary adipose tissue on non-contrast computerized tomography for predicting non-calcified plaques

Supplemental material, sj-pdf-1-xst-10.1177_08953996241292476 for Radiomics and deep learning features of pericoronary adipose tissue on non-contrast computerized tomography for predicting non-calcified plaques by Junli Yu, Yan Ding, Li Wang, Shunxin Hu, Ning Dong, Jiangnan Sheng, Yingna Ren and Ziyue Wang in Journal of X-Ray Science and Technology

Footnotes

Acknowledgements

The authors would like to thank Prof. Li Wang and Prof. Yan Ding for their guidance on my experiment, as well as Ning Dong and Shunxin Hu for their assistance with data analysis. We also appreciate the help of Jiangnan Sheng, Yingna Ren, and Ziyue Wang in collecting the cases.

Ethics approval and consent to participate

The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). The study was approved by the Institutional Review Board (IRB) of Qiqihar Medical University, and individual consent for this retrospective analysis was waived.

Consent for publication

Consent forms were obtained.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work is supported by the Qiqihar Medical University Postgraduate Innovation Fund Project (Grant number QYYCX2023-11).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available due confidential information but are available from the corresponding author on reasonable request.*

Supplemental material

Supplemental material for this article is available online.

References

Vaduganathan

Mensah

Turco

, et al. The global burden of cardiovascular diseases and risk: a compass for future health. J Am Coll Cardiol 2022; 80: 2361–2371.

Libby

Hansson

. From focal lipid storage to systemic inflammation: JACC review topic of the week. J Am Coll Cardiol 2019; 74: 1594–1607.

Antonopoulos

Sanna

Sabharwal

, et al. Detecting human coronary inflammation by imaging perivascular fat. Sci Transl Med 2017; 9: eaal2658.

Antonopoulos

Margaritis

Coutinho

, et al. Reciprocal effects of systemic inflammation and brain natriuretic peptide on adiponectin biosynthesis in adipose tissue of patients with ischemic heart disease. Arterioscler Thromb Vasc Biol 2014; 34: 2151–2159.

Antonopoulos

Margaritis

Coutinho

, et al. Adiponectin as a link between type 2 diabetes and vascular NADPH oxidase activity in the human arterial wall: the regulatory role of perivascular adipose tissue. Diabetes 2015; 64: 2207–2219.

Margaritis

Antonopoulos

Digby

, et al. Interactions between vascular wall and perivascular adipose tissue reveal novel roles for adiponectin in the regulation of endothelial nitric oxide synthase function in human vessels. Circulation 2013; 127: 2209–2221.

Oikonomou

Marwan

Desai

, et al. Non-invasive detection of coronary inflammation using computed tomography and prediction of residual cardiovascular risk (the CRISP CT study): a post-hoc analysis of prospective outcome data. Lancet 2018; 392: 929–939.

Antoniades

Antonopoulos

Deanfield

. Imaging residual inflammatory cardiovascular risk. Eur Heart J 2020; 41: 748–758.

American College of Cardiology Foundation Task Force on Expert Consensus Documents; Mark

Berman

Budoff

, et al. ACCF/ACR/AHA/NASCI/SAIP/SCAI/SCCT 2010 expert consensus document on coronary computed tomographic angiography: a report of the American College of Cardiology Foundation Task Force on Expert Consensus Documents. Circulation 2010; 121: 2509–2543.

10.

Almeida

Pelter

Shaikh

, et al. Feasibility of measuring pericoronary fat from precontrast scans: effect of iodinated contrast on pericoronary fat attenuation. J Cardiovasc Comput Tomogr 2020; 14: 490–494.

11.

Plank

Friedrich

Dichtl

, et al. The diagnostic and prognostic value of coronary CT angiography in asymptomatic high-risk patients: a cohort study. Open Heart 2014; 1: e000096.

12.

Cury

Abbara

Achenbach

, et al. CAD-RADS(TM) Coronary Artery Disease - Reporting and Data System. An expert consensus document of the Society of Cardiovascular Computed Tomography (SCCT), the American College of Radiology (ACR) and the North American Society for Cardiovascular Imaging (NASCI). Endorsed by the American College of Cardiology. J Cardiovasc Comput Tomogr 2016; 10: 269–281.

13.

Kirişli

Schaap

Metz

, et al. Standardized evaluation framework for evaluating coronary artery stenosis detection, stenosis quantification and lumen segmentation algorithms in computed tomography angiography. Med Image Anal 2013; 17: 859–876.

14.

Arbab-Zadeh

Hoe

. Quantification of coronary arterial stenoses by multidetector CT angiography in comparison with conventional angiography methods, caveats, and implications. JACC Cardiovasc Imaging 2011; 4: 191–202.

15.

Taron

Foldyna

Mayrhofer

, et al. Risk stratification with the use of coronary computed tomographic angiography in patients with nonobstructive coronary artery disease. JACC Cardiovasc Imaging 2021; 14: 2186–2195.

16.

Sohn

Fields

BKK

. Radiomics and deep learning to predict pulmonary nodule metastasis at CT. Radiology 2024; 311: e233356.

17.

Morales

Manning

Nezafat

. Present and future innovations in AI and cardiac MRI. Radiology 2024; 310: e231269.

18.

Fan

You

, et al. Dual-energy CT deep learning radiomics to predict macrotrabecular-massive hepatocellular carcinoma. Radiology 2023; 308: e230255.

19.

Goeller

Tamarappoo

Kwan

, et al. Relationship between changes in pericoronary adipose tissue attenuation and coronary plaque burden quantified from coronary computed tomography angiography. Eur Heart J Cardiovasc Imaging 2019; 20: 636–643.

20.

Elnabawi

Oikonomou

Dey

, et al. Association of biologic therapy with coronary inflammation in patients with psoriasis as assessed by perivascular fat attenuation index. JAMA Cardiol 2019; 4: 885–891.

21.

Zhang

Gao

Zhang

, et al. Personalized CT-based radiomics nomogram preoperative predicting Ki-67 expression in gastrointestinal stromal tumors: a multicenter development and validation cohort. Clin Transl Med 2020; 9: 12.

22.

Zhang

Ren

, et al. Deep residual learning for image recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016.

23.

Zhang

Gao

Zhang

24.

Lundberg

Lee

SIJ

. A unified approach to interpreting model predictions. arXiv preprint 2017; 30: 1705.07874.

25.

Athanasiou

Sfrintzeri

Zarkogianni

, et al. An explainable XGBoost-based approach towards assessing the risk of cardiovascular disease in patients with Type 2 Diabetes Mellitus. In: 2020 IEEE 20th International Conference on Bioinformatics and Bioengineering BIBE); 2020, IEEE.

26.

Lundberg

Erion

Lee

SIJ

. Consistent individualized feature attribution for tree ensembles. arXiv preprint 2018; 1802.03888.

27.

Lundberg

Nair

Vavilala

, et al. Explainable machine-learning predictions for the prevention of hypoxaemia during surgery. Nat Biomed Eng 2018; 2: 749–760.

28.

Vickers

Holland

. Decision curve analysis to evaluate the clinical benefit of prediction models. Spine J 2021; 21: 1643–1648.

29.

DeLong

Clarke-Pearson

. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics 1988; 44: 837–845.

30.

Timmis

Townsend

Gale

, et al. European society of cardiology: cardiovascular disease statistics 2017. Eur Heart J 2018; 39: 508–579.

31.

Kolossváry

Karády

Kikuchi

, et al. Radiomics versus visual and histogram-based assessment to identify atheromatous lesions at coronary CT angiography: an ex vivo study. Radiology 2019; 293: 89–96.

32.

Zreik

Lessmann

Hamersvelt

, et al. Deep learning analysis of the myocardium in coronary CT angiography for identification of patients with functionally significant coronary artery stenosis. Med Image Anal 2018; 44: 72–85.

33.

Homayounieh

Yan

Digumarthy

, et al. Prediction of coronary calcification and stenosis: role of radiomics from low-dose CT. Acad Radiol 2021; 28: 972–979.

34.

Tao

Wang

, et al. Feasibility on the diagnosis of non-calcified plaque based on radiomics of pericoronary adipose tissue on plain CT scan image. Zhonghua Yi Xue Za Zhi 2021; 101: 458–463. Chinese.

35.

Jiang

Shao

Chai

, et al. Non-contrast CT-based radiomic signature of pericoronary adipose tissue for screening non-calcified plaque. Phys Med Biol 2022; 67: 10.

36.

Xie

Yang

, et al. Computed tomography-based deep-learning prediction of neoadjuvant chemoradiotherapy treatment response in esophageal squamous cell carcinoma. Radiother Oncol 2021; 154: 6–13.

37.

Falk

. Pathogenesis of atherosclerosis. J Am Coll Cardiol 2006; 47: C7–12.

38.

Lin

Kolossváry

Cadet

, et al. Radiomics-Based precision phenotyping identifies unstable coronary plaques from computed tomography angiography. JACC Cardiovasc Imaging 2022; 15: 859–871.

39.

Tomaszewski

Gillies

. The biological meaning of radiomic features. Radiology 2021; 298: 505–516.

40.

Homayounieh

Yan

Digumarthy

, et al. Prediction of coronary calcification and stenosis: role of radiomics from low-dose CT. Acad Radiol 2021; 28: 972–979.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

1.00 MB