Skip to main content
NCBI home page
Frontiers in Endocrinology logoLink to Frontiers in Endocrinology
. 2026 Aug 7;17:1908048. doi: 10.3389/fendo.2026.1908048

AI-assisted segmentation-based fusion model integrating deep learning, radiomics, and clinical parameters for identifying coronary heart disease risk in patients with MAFLD

Luying Qi 1,†, Shumin Zhang 2,†, Bin Wang 3, Mei Yang 1, Yuanbin Li 1, Baiwang Li 2, Jingnan Xue 1, Xuemei Li 3,*
PMCID: PMC13504196  PMID: 42643589

Abstract

Background/objectives

Metabolic dysfunction-associated fatty liver disease (MAFLD) is associated with an increased risk of coronary heart disease (CHD), which is one of the leading causes of chronic disease-related mortality worldwide. Early identification of CHD risk in patients with MAFLD is essential for risk stratification and timely intervention. Although radiomics and deep learning (DL) have shown promising performance in medical image analysis, their application for CHD risk assessment in patients with MAFLD remains limited. Therefore, this study aimed to develop and validate an AI-assisted segmentation-based deep learning radiomics-clinical (DLRC) model for identifying CHD risk in patients with MAFLD.

Methods

A total of 1,515 patients with MAFLD, including 440 with concomitant CHD, were retrospectively enrolled between January 2023 and December 2025. Patients were randomly divided into a training cohort (n = 1,060) and a test cohort (n = 455). AI-assisted liver segmentation was performed using a deep learning-assisted framework integrated into ITK-SNAP. Radiomics features were extracted using PyRadiomics, and DL features were extracted using a pre-trained DenseNet-121 network. After feature selection using Pearson correlation analysis, minimum redundancy maximum relevance (mRMR), principal component analysis (PCA), and least absolute shrinkage and selection operator (LASSO) regression, radiomics and DL features were integrated to construct a deep learning radiomics (DLR) model. Clinical predictors were identified using multivariable logistic regression. Clinical, radiomics, DLR, and combined DLRC models were developed and evaluated using machine learning algorithms. Model performance was assessed by receiver operating characteristic (ROC) analysis, calibration curves, DeLong tests, and decision curve analysis (DCA).

Results

Multivariable analysis identified older age, hypertension, diabetes mellitus, hyperlipidemia, male sex, and lower BMI as independent predictors of CHD in patients with MAFLD. Eleven radiomics features and fifteen DLR features were selected for model construction. The DLRC model achieved the best predictive performance, with AUCs of 0.917 in the training cohort and 0.878 in the test cohort, outperforming the DLR model (0.895 and 0.854), radiomics model (0.821 and 0.784), clinical model (0.768 and 0.751), respectively. DeLong tests demonstrated significant superiority of the DLRC model over the other models (all P < 0.05). Calibration and DCA analyses further confirmed its excellent calibration and clinical utility.

Conclusion

An AI-assisted segmentation-based fusion model integrating clinical factors, radiomics features, and DL features demonstrated excellent performance for identifying CHD risk in patients with MAFLD. This non-invasive and efficient approach may facilitate early risk stratification and personalized management of MAFLD patients.

Keywords: artificial intelligence, clinical parameters, coronary heart disease, deep learning, metabolic dysfunction-associated fatty liver disease, radiomics, risk prediction

1. Introduction

In 2020, an international expert consensus statement proposed replacing non-alcoholic fatty liver disease (NAFLD) with metabolic dysfunction-associated fatty liver disease (MAFLD), aiming to eliminate the potential stigmatization associated with the terms “fatty” and “alcoholic” in NAFLD, while adopting affirmative diagnostic criteria instead of the previous exclusionary criteria (1). Currently, approximately one-third (32.4%) of the global adult population is affected by MAFLD, making it the most common chronic liver disease worldwide (2). MAFLD is closely associated with various systemic disorders, not only impairing patients’ health-related quality of life but also increasing healthcare costs and economic burden (3). Multiple studies have shown that MAFLD is significantly associated with an increased risk of cardiovascular disease (CVD) (4–7); however, the definition of cardiovascular outcomes in these studies is relatively broad. Coronary heart disease (CHD) is the most common type of cardiovascular disease and a leading cause of chronic disease-related mortality worldwide (8). In recent investigations, MAFLD has been demonstrated to be associated with an increased risk of CHD (9). The specific mechanisms underlying the association between MAFLD and CHD remain unclear, with genetic susceptibility, insulin resistance (IR), dyslipidemia, oxidative stress, inflammation, and endothelial dysfunction considered key links mediating the increased CHD risk in NAFLD (10). Therefore, early detection of MAFLD patients and timely intervention are of paramount importance.

Artificial intelligence (AI) has achieved significant progress in the fields of disease diagnosis and medical image analysis (11, 12). Radiomics employs high-throughput automated extraction algorithms to evaluate features such as geometric shape and texture that are imperceptible to the human visual system, transforming traditional images that rely on visual interpretation into quantitative features (13). Deep learning (DL), as an important branch of AI, is particularly adept at identifying complex features in radiological images. Through multi-layer neural network (NN) models and utilizing large-scale data, DL achieves efficient learning and feature extraction, thereby improving the accuracy of disease diagnosis and assessment (11, 14). Currently, radiomics and DL have been extensively studied in malignant tumors, including lung cancer, colorectal cancer, gastric cancer, and breast cancer (15–18); however, their application in chronic diseases remains relatively limited.

To date, no study has utilized AI-assisted segmentation-based DL radiomics to predict CHD risk in patients with MAFLD. In the present study, we integrated DL features, radiomics features, and clinical variables into a final predictive model and evaluated its diagnostic performance for identifying CHD in patients with MAFLD.

2. Methods

This retrospective study was designed to develop and validate an AI-assisted segmentation-based DLRC model for CHD risk identification in patients with MAFLD. The overall workflow consisted of five major steps: patient selection, liver CT preprocessing and AI-assisted segmentation, extraction of radiomics and deep learning features, integration with clinical parameters, and model evaluation.

2.1. Patient selection and clinical data

This retrospective study was approved by the Ethics Committee of First Affiliated Hospital of Huzhou Normal University (Approval Number: 2025KYLL099-01), which waived the requirement for informed consent. Clinical data and CT images were collected from 1,515 eligible patients (1,075: MAFLD without CHD and 440: MAFLD with CHD) between January 2023 and December 2025. Inclusion criteria were: (1) age ≥ 18 years; (2) diagnosis of MAFLD according to the Primary Care Guidelines for the Diagnosis and Management of Metabolic Dysfunction-Associated Fatty Liver Disease (2025) (19); (3) availability of complete thin-slice (1-mm) liver CT images; and (4) liver CT performed within one week of MAFLD diagnosis. Exclusion criteria were: (1) incomplete clinical data; (2) alcohol consumption ≥ 210 g/week in male patients or ≥ 140 g/week in female patients; (3) presence of specific factors leading to fatty liver, including chronic viral hepatitis, hereditary metabolic diseases (e.g., Wilson’s disease), drug-induced liver disease, total parenteral nutrition, or a history of liver resection; (4) other cardiac diseases (e.g., myocarditis, cardiomyopathy, valvular heart disease) or history of CABG/PCI; (5) history of malignancy; or (6) soft plaques on coronary angiography/CTA. The 1,515 patients were randomly divided into a training cohort (n = 1,060) and a test cohort (n = 455). Figure 1 shows the patient selection flowchart.

Figure 1.

Flowchart describes patient selection for a study on MAFLD and coronary heart disease, showing inclusion and exclusion criteria, resulting in 1515 eligible patients, then split into training (1060) and test (455) cohorts.

Open in a new tab

Flowchart of patient selection.

Clinical information included sex, age, body mass index (BMI), hypertension, diabetes mellitus, hyperlipidemia, smoking status, and hepatic insufficiency status. Laboratory indicators included white blood cell count (WBC), neutrophil count (NEU), absolute lymphocyte count (LYM), absolute monocyte count (MONO), hemoglobin (Hb), platelet count (PLT), albumin (ALB), glucose (Glu), triglycerides (TG), total cholesterol (TC), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), aspartate aminotransferase (AST), alanine aminotransferase (ALT), total bilirubin (TBIL), creatinine (Cr), blood urea nitrogen (BUN), uric acid (UA), and the albumin-to-globulin ratio (A/G).

2.2. Image acquisition and preprocessing

Patients were instructed to hold their breath after calm inspiration, and non-contrast computed tomography (CT) of the liver was performed. Each patient was examined using one of the following scanners: Somatom Definition AS (Siemens Healthineers), Somatom Perspective (Siemens Healthineers), or GE Optima 540 (GE Healthcare). Detailed scanning parameters are provided in Supplementary. All images were standardized through a three-step preprocessing procedure. First, all images were converted from DCM format to nii.gz format. Second, images were resampled to a voxel size of 1 × 1 × 1 mm³. Finally, a uniform window width of 400 Hounsfield units (HU) and a window level of 70 HU were applied to all images.

2.3. AI-assisted liver image segmentation

AI-assisted liver segmentation was performed using the deep learning server (DLS) within ITK-SNAP (version 4.4), a Python-based extension that integrates nnInteractive’s deep learning interactive segmentation functions into the ITK-SNAP platform. After simple annotation with a brush tool, the system automatically generated an accurate 3D liver segmentation within seconds, significantly reducing manual workload and improving efficiency. All imaging data were anonymized before evaluation, and any patient identifiers were removed. All generated liver masks were visually inspected to ensure anatomical accuracy, and necessary manual corrections were performed by a senior radiologist when discrepancies were identified, particularly in regions with unclear liver boundaries, adjacent organs, or heterogeneous liver parenchyma. The entire liver volume was then extracted as the region of interest (ROI) and used as input for the subsequent 3D deep learning model.

2.4. Feature extraction and model development

Radiomics features were extracted using the PyRadiomics package within the OnekeyAI platform. All features were normalized using Z-score normalization to eliminate discrepancies caused by differences in units, scales, and imaging equipment across features. To reduce multicollinearity, features with a Pearson correlation coefficient > 0.9 were iteratively excluded. The minimum redundancy maximum relevance (mRMR) method was then applied to remove irrelevant and redundant features. Subsequently, LASSO regression was employed to reduce the dimensionality of the initial features, and weights were assigned to the selected features. Features with higher weights were retained and used as inputs for machine learning classifiers in model development. To minimize overfitting, feature selection procedures were performed exclusively within the training cohort, and five-fold cross-validation was used to determine the optimal LASSO penalty parameter before final model construction.

DL features were extracted using the ImageNet-pretrained DenseNet-121 model via transfer learning. DenseNet-121 is a classic convolutional neural network (CNN) architecture, and was employed to extract DL features in this study. The extracted DL features were compressed using principal component analysis (PCA) and subsequently normalized using Z-score standardization. LASSO regression was then applied to filter features with non-zero coefficients, thereby reducing dimensionality and selecting the compressed DL features. Radiomics features and compressed DL features were integrated to develop deep learning radiomics (DLR) features. After LASSO-based feature selection, the resulting DLR features were used as inputs for model development.

For clinical features, univariable and multivariable logistic regression analyses were performed. Variables with a P-value < 0.05 in univariable analysis were entered into the multivariable analysis to identify statistically significant features for constructing the clinical model. The clinical features, radiomics features, and DLR features were separately fed into eight machine learning classifiers, including logistic regression (LR), support vector machine (SVM), K-nearest neighbors (KNN), random forest (RF), ExtraTrees, gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and multilayer perceptron (MLP), to develop models. The optimal models based on clinical, radiomics, and DLR features were selected respectively. The clinical model and the DLR model were subsequently integrated to create the final predictive model in this study, which was termed the DLRC model (deep learning-radiomics-clinical model). Model performance was evaluated using the following metrics: accuracy, area under the curve (AUC) with 95% confidence interval (CI), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), precision, recall, and F1-score. Calibration curves were used to assess the calibration performance of the combined model in both the training and test datasets, and the Hosmer–Lemeshow test was applied to evaluate goodness-of-fit. Decision curve analysis (DCA) was performed to assess the net benefit of the model across different probability thresholds. The study design workflow is illustrated in Figure 2.

Figure 2.

Flowchart illustrating a machine learning pipeline for liver segmentation using CT images, including sections for image input, segmentation, DenseNet121-based feature extraction, radiomics features, feature selection with LASSO and MSE plots, and model evaluation with AUC, calibration, and DCA curve charts.

Open in a new tab

Flowchart of the study design.

2.5. Statistical analysis

Statistical analysis of clinical variables was performed using SPSS (version 26.0). Continuous variables with a normal distribution were compared using the independent samples t-test, while those with a non-normal distribution were compared using the Mann–Whitney U test. Categorical variables are presented as n (%) and were compared using the chi-square test. Baseline comparisons between CHD and non-CHD groups were performed using independent samples t-tests, Mann–Whitney U test or chi-square tests. Univariable and multivariable binary logistic regression analyses were performed to identify independent clinical risk factors associated with CHD in patients with MAFLD. Variables with statistical significance in univariable analyses were subsequently entered into multivariable logistic regression analysis. A P-value < 0.05 was considered statistically significant. Feature extraction and modeling analyses for radiomics, deep learning, and clinical variables were all performed using the Python environment on the OnekeyAI platform (version 4.9.1).

Based on the above workflow, clinical characteristics, imaging features, DLR features, and integrated predictive models were subsequently analyzed and compared between patients with and without CHD.

3. Results

3.1. Clinical characteristics and univariate and multivariate analyses

As of December 2025, a total of 1,515 eligible patients with MAFLD were enrolled and randomly divided into a training cohort (n = 1,060; 749 with MAFLD without CHD and 311 with MAFLD with CHD) and a test cohort (n = 455; 326 with MAFLD without CHD and 129 with MAFLD with CHD) at a ratio of 7:3. The study population comprised 903 males and 612 females. In both the training and test cohorts, statistically significant differences were observed between the non-CHD and CHD groups in terms of age, ALB, TC, LDL-C, Cr, BUN, and hypertension (all P < 0.05). Additionally, in the training cohort, significant differences were also noted between the two groups in WBC, NEU, MONO, Hb, HDL-C, ALT, A/G, diabetes, and hepatic insufficiency (all P < 0.05). Detailed information is presented in Table 1.

Table 1.

Baseline characteristics of MAFLD patients with and without CHD in the training and test cohorts.

Clinical factors Training cohort (n=1060) P value Test cohort (n=455) P value
MAFLD without CHD (n=749) MAFLD with CHD (n=311) MAFLD without CHD (n=326) MAFLD with CHD (n=129)
Age (years) 51.54 ± 15.40 63.74 ± 10.66 <0.001 52.84 ± 15.58 63.62 ± 11.18 <0.001
BMI (kg/m²) 27.04 ± 4.44 26.57 ± 3.64 0.292 26.83 ± 3.95 26.82 ± 3.26 0.564
WBC (×109/L) 8.14 ± 3.68 6.61 ± 2.03 <0.001 7.59 ± 3.36 6.91 ± 2.61 0.124
NEU (×109/L) 5.92 ± 3.61 4.37 ± 1.79 <0.001 5.42 ± 3.21 5.02 ± 4.87 0.128
LYM (×109/L) 1.63 ± 0.83 1.68 ± 0.64 0.055 1.63 ± 0.80 1.72 ± 0.66 0.045
MONO (×109/L) 0.46 ± 0.24 0.39 ± 0.17 <0.001 0.42 ± 0.22 0.40 ± 0.15 0.997
Hb (g/L) 140.99 ± 20.75 143.58 ± 17.11 0.038 141.08 ± 19.32 142.90 ± 18.09 0.223
PLT (×109/L) 219.16 ± 76.23 209.80 ± 58.36 0.099 217.03 ± 70.18 205.33 ± 59.90 0.110
ALB (g/L) 41.73 ± 5.00 43.26 ± 3.72 <0.001 42.30 ± 4.34 43.34 ± 3.46 0.016
Glu (mmol/L) 7.63 ± 3.68 7.61 ± 3.42 0.529 7.56 ± 3.48 7.28 ± 2.92 0.853
TG (mmol/L) 3.71 ± 7.05 2.29 ± 1.58 0.758 3.84 ± 7.24 2.43 ± 1.28 0.597
TC (mmol/L) 5.29 ± 1.95 4.68 ± 1.36 <0.001 5.42 ± 2.14 4.63 ± 1.13 <0.001
HDL-C (mmol/L) 74.95 ± 33.59 81.08 ± 29.96 0.010 78.43 ± 31.95 76.03 ± 25.83 0.455
LDL-C (mmol/L) 2.73 ± 0.93 2.52 ± 0.97 0.002 2.76 ± 0.92 2.52 ± 0.89 0.026
AST (U/L) 36.19 ± 49.81 31.06 ± 25.06 0.446 40.36 ± 53.49 30.07 ± 15.73 0.131
ALT (U/L) 47.35 ± 63.41 37.03 ± 28.87 0.015 50.05 ± 70.29 37.09 ± 25.16 0.073
TBIL (μmol/L) 17.29 ± 12.58 16.09 ± 10.30 0.580 17.50 ± 13.31 15.94 ± 8.05 0.474
Cr (μmol/L) 77.92 ± 53.42 79.24 ± 27.44 <0.001 76.26 ± 26.95 86.12 ± 55.52 0.012
BUN (mmol/L) 5.46 ± 2.71 5.99 ± 2.23 <0.001 5.49 ± 2.10 6.24 ± 2.03 <0.001
UA (μmol/L) 380.01 ± 135.07 385.56 ± 104.78 0.169 380.26 ± 112.38 383.71 ± 105.91 0.355
A/G 1.48 ± 0.27 1.53 ± 0.24 0.004 1.48 ± 0.27 1.51 ± 0.22 0.295
Sex 0.712 0.061
Female 295(39.39) 118(37.94) 152(46.63) 47(36.43)
Male 454(60.61) 193(62.06) 174(53.37) 82(63.57)
Hypertension <0.001 <0.001
No 401(53.54) 69(22.19) 159(48.77) 25(19.38)
Yes 348(46.46) 242(77.81) 167(51.23) 104(80.62)
Diabetes <0.001 0.573
No 522(69.69) 180(57.88) 223(68.40) 84(65.12)
Yes 227(30.31) 131(42.12) 103(31.60) 45(34.88)
Hyperlipidemia 0.920 0.954
No 579(77.30) 242(77.81) 256(78.53) 101(78.29)
Yes 170(22.70) 69(22.19) 70(21.47) 28(21.71)
Smoking 0.457 0.943
No 590(78.77) 252(81.03) 255(78.22) 102(79.07)
Yes 159(21.23) 59(18.97) 71(21.78) 27(20.93)
Hepatic Insufficiency 0.022 0.213
No 624(83.31) 277(89.07) 274(84.05) 115(89.15)
Yes 125(16.69) 34(10.93) 52(15.95) 14(10.85)
Open in a new tab

To identify independent clinical predictors associated with CHD, univariable binary logistic regression analyses were first performed in the training cohort. Variables were then entered into multivariable logistic regression analysis when P values < 0.05. After adjustment for confounders, hyperlipidemia, male sex, diabetes, hypertension, lower BMI, and older age remained independently associated with CHD (all P < 0.05). Variance inflation factor (VIF) values were < 5 for all variables, indicating no significant collinearity. The OR (95% CI) were 1.949 (1.376–2.759), 1.600 (1.124–2.277), 1.447 (1.064–1.970), 2.541 (1.896–3.408), 0.932 (0.900–0.965), and 1.040 (1.029–1.051), respectively. Hyperlipidemia, male sex, diabetes, hypertension, lower BMI, and older age are associated with higher CHD risk in MAFLD patients, with hypertension being the strongest predictor. These variables were used to develop the clinical model (Table 2).

Table 2.

Univariable and multivariable logistic regression analyses of clinical factors associated with CHD.

Clinical factors Univariable analysis Multivariable analysis
OR (95% CI) P value OR (95% CI) P value
Age (years) 0.989 (0.987–0.991) <0.001 1.040 (1.029–1.051) <0.001
BMI (kg/m²) 0.968 (0.964–0.971) <0.001 0.932 (0.900–0.965) 0.001
WBC (×109/L) 0.881 (0.868–0.895) <0.001 1.065 (0.749–1.514) 0.769
NEU (×109/L) 0.835 (0.817–0.853) <0.001 0.784 (0.549–1.120) 0.261
LYM (×109/L) 0.638 (0.598–0.680) <0.001 1.033 (0.692–1.542) 0.894
MONO (×109/L) 0.133 (0.103–0.171) <0.001 0.319 (0.122–0.838) 0.052
Hb (g/L) 0.994 (0.993–0.995) <0.001 1.006 (0.997–1.015) 0.279
PLT (×109/L) 0.996 (0.996–0.997) <0.001 1.001 (0.999–1.004) 0.295
ALB (g/L) 0.980 (0.978–0.983) <0.001 0.999 (0.960–1.039) 0.952
Glu (mmol/L) 0.905 (0.892–0.918) <0.001 0.990 (0.946–1.037) 0.715
TG (mmol/L) 0.763 (0.731–0.796) <0.001 0.958 (0.861–1.067) 0.518
TC (mmol/L) 0.834 (0.815–0.852) <0.001 0.775 (0.593–1.012) 0.116
HDL-C (mmol/L) 0.991 (0.990–0.992) <0.001 1.005 (1.000–1.011) 0.127
LDL-C (mmol/L) 0.722 (0.694–0.753) <0.001 0.863 (0.629–1.184) 0.442
AST (U/L) 0.978 (0.975–0.981) <0.001 1.000 (0.990–1.009) 0.924
ALT (U/L) 0.982 (0.979–0.985) <0.001 0.998 (0.990–1.005) 0.615
TBIL (μmol/L) 0.954 (0.948–0.961) <0.001 0.985 (0.969–1.001) 0.119
Cr (μmol/L) 0.990 (0.989–0.991) <0.001 0.995 (0.988–1.001) 0.142
BUN (mmol/L) 0.883 (0.867–0.900) <0.001 1.003 (0.932–1.078) 0.955
UA (μmol/L) 0.998 (0.998–0.998) <0.001 1.001 (1.000–1.003) 0.165
A/G 0.578 (0.537–0.622) <0.001 0.671 (0.362–1.241) 0.286
Sex 0.425 (0.369–0.490) <0.001 1.600 (1.124–2.277) 0.029
Hypertension 0.695 (0.606–0.798) <0.001 2.541 (1.896–3.408) <0.001
Diabetes 0.577 (0.482–0.691) <0.001 1.447 (1.064–1.970) 0.048
Hyperlipidemia 0.406 (0.321–0.513) <0.001 1.949 (1.376–2.759) 0.002
Smoking 0.371 (0.289–0.477) <0.001 1.101 (0.773–1.570) 0.655
Hepatic Insufficiency 0.272 (0.198–0.374) <0.001 0.827 (0.532–1.287) 0.479
Open in a new tab

3.2. Feature selection and model development

A total of 107 radiomics features were extracted. After Z-score normalization, Pearson correlation, mRMR, and LASSO regression with five-fold cross-validation, the optimal λ value was determined as 0.0072, yielding 11 radiomics features with non-zero coefficients. These features were used to construct the radiomics model. Subsequently, 31 compressed DL features were integrated with the radiomics features. Following Pearson correlation, LASSO regression, and five-fold cross-validation (λ = 0.0041), 15 DLR features with non-zero coefficients were selected to construct the DLR model. The Radscore and DLRscore are provided in Supplementary. Visualization of LASSO coefficient paths, MSE curves, and feature weights for both radiomics and DLR features is presented in Figure 3.

Figure 3.

Six-panel figure showing results of LASSO regression analyses: (a, d) line plots of coefficient trajectories versus Lambda; (b, e) mean squared error plots with cross-validation and error bars; (c, f) horizontal bar charts of top feature coefficients, with feature names on the y-axis and coefficient values on the x-axis. Upper panels use conventional radiomic features and lower panels use a combination of deep learning and radiomic features.

Open in a new tab

(a, d) LASSO coefficient paths for radiomics and DLR features; (b, e) MSE curves for radionics and DLR features; (c, f) Feature weights for radionics and DLR features.

The DLR features were combined with the selected clinical features to construct a fusion model, which we termed the DLRC model. The nomogram of this model, which can predict the probability of CHD based on multiple indicators, is shown in Figure 4. Our nomogram demonstrated that the DLR score was the most important predictor, followed by age.

Figure 4.

Nomogram graphic displays scales for calculating risk based on variables including BMI, age, diabetes, sex, hyperlipidemia, hypertension, and DLR, with total points corresponding to estimated risk values.

Open in a new tab

Nomogram of the DLRC model.

Table 3 summarizes the performance of the four models. In the training cohort, the fusion model achieved the highest AUC (0.917), followed by the DLR model (0.895), radiomics model (0.821), and clinical model (0.768). Similar trends were observed in the test cohort, with AUCs of 0.878, 0.854, 0.784, and 0.751, respectively. The fusion model consistently outperformed the other models in both cohorts. The receiver operating characteristic (ROC) curves are displayed in Figure 5. DeLong tests confirmed that the combined model significantly outperformed the clinical, radiomics, and DLR models in both cohorts (all P < 0.05), as shown in Figure 6.

Table 3.

Performance comparison of the four models in the training and test cohorts.

Signature Accuracy AUC(95% CI) Sensitivity Specificity PPV NPV Precision Recall F1 Threshold Cohort
Clinic 0.659 0.768 (0.739–0.797) 0.862 0.575 0.457 0.909 0.457 0.862 0.598 0.222 train
Rad 0.717 0.821 (0.795–0.848) 0.836 0.668 0.511 0.907 0.511 0.836 0.634 0.283 train
DLR 0.851 0.895 (0.873–0.918) 0.762 0.888 0.738 0.900 0.738 0.762 0.750 0.370 train
Combined 0.861 0.917 (0.899–0.935) 0.794 0.889 0.748 0.912 0.748 0.794 0.771 0.381 train
Clinic 0.648 0.751 (0.706–0.797) 0.845 0.571 0.438 0.903 0.438 0.845 0.577 0.240 test
Rad 0.705 0.784 (0.738–0.830) 0.798 0.669 0.488 0.893 0.488 0.798 0.606 0.284 test
DLR 0.793 0.854 (0.813–0.895) 0.760 0.807 0.609 0.895 0.609 0.760 0.676 0.323 test
Combined 0.787 0.878 (0.844–0.912) 0.845 0.764 0.586 0.926 0.586 0.845 0.692 0.202 test
Open in a new tab

Figure 5.

Two side-by-side ROC curve charts labeled a and b compare four predictive models using sensitivity versus one minus specificity. In both charts, combined models (orange dashed lines) achieve the highest AUC, followed by DLR (cyan), Rad (navy dashed), and Clinic (magenta). Chart a shows training cohort results, with combined AUC at 0.917, DLR at 0.895, Rad at 0.821, and Clinic at 0.768. Chart b shows test cohort results, with combined AUC at 0.878, DLR at 0.854, Rad at 0.784, and Clinic at 0.751. Legends and confidence intervals are included for each curve.

Open in a new tab

ROC curves of different models (a) training cohort; (b) testing cohort.

Figure 6.

Two color-coded heatmaps labeled a and b display Delong test results for train and test cohorts, respectively. Rows and columns categorize Clinic, Rad, DLR, and Combined models, with values annotated within each square. Color bars indicate value scales from red to blue; heatmap a ranges up to 0.0022 and heatmap b up to 0.25. Higher values are predominantly in upper left squares, while most other entries are zero.

Open in a new tab

DeLong test results for comparing the AUCs of different models (a) training cohort; (b) testing cohort.

3.3. Comparison of clinical utility among different models

Calibration curves are shown in Figure 7. The results demonstrated that the actual reference lines of all models were in good agreement with the ideal reference lines, with the fusion model exhibiting the best calibration performance. DCA was further performed to evaluate the clinical utility of the models. The results showed that the fusion model achieved a wider range of net benefit across threshold probabilities compared to the other models. Moreover, within most threshold probability ranges, the fusion model yielded higher net benefits than the DLR, radiomics, and clinical models (Figure 8).

Figure 7.

Two side-by-side line graphs labeled a and b display calibration plots for cohort train and cohort test data, respectively. Each graph compares four model types—Clinic, Rad, DLR, and Combined—against a dotted line representing perfect calibration. The plots show the relationship between mean predicted probability and the fraction of positives for each model, with all lines generally following but deviating from the diagonal perfect calibration line. Legends and axes are clearly marked.

Open in a new tab

Calibration curves of different models (a) training cohort; (b) testing cohort.

Figure 8.

Figure contains two side-by-side line graphs labeled a and b, illustrating decision curve analyses for train and test cohorts. Both graphs plot net benefit against threshold probability, comparing Clinic, Rad, DLR, and Combined methods, along with Treat all and Treat none reference lines. Combined method (red) shows the highest net benefit across most threshold probabilities in both panels.

Open in a new tab

Decision curve analysis of different models (a) training cohort; (b) testing cohort.

3.4. Subgroup analysis

Further subgroup analyses were performed in the test cohort to evaluate the robustness of the combined model. The same classification threshold was applied across all subgroups to ensure objective comparison and avoid subgroup-specific optimization, which may introduce overfitting. The results demonstrated that the combined model maintained good discriminatory ability across different sex and age subgroups. The details are shown in Figure 9 and Table 4.

Figure 9.

Four Receiver Operating Characteristic (ROC) curve graphs compare sensitivity versus one minus specificity for different groups: a, females; b, males; c, age less than sixty years; d, age sixty years or older. Each chart features a blue ROC curve and a red diagonal reference line.

Open in a new tab

ROC curves of the combined model in different sex and age subgroups (a) females; (b) males; (c) patients aged <60 years; (d) patients aged ≥60 years.

Table 4.

Subgroup analysis of the combined model performance in the test cohort.

Subgroup n AUC 95%CI Sensitivity Specificity
Sex Female 199 0.863 0.801-0.925 0.872 0.743
Male 256 0.891 0.852-0.931 0.829 0.782
Age (years) Age < 60 256 0.935 0.894-0.977 0.826 0.914
Age ≥ 60 199 0.760 0.692-0.828 0.855 0.491
Open in a new tab

In the sex subgroup analysis, the combined model achieved an AUC of 0.863 (95% CI: 0.801–0.925) in female patients and 0.891 (95% CI: 0.852–0.931) in male patients, indicating comparable predictive performance between the two groups. The sensitivity and specificity were 0.872 and 0.743 in females and 0.829 and 0.782 in males, respectively. In the age subgroup analysis, the combined model showed superior performance in patients aged <60 years, with an AUC of 0.935 (95% CI: 0.894–0.977), sensitivity of 0.826, and specificity of 0.914. The model demonstrated moderate discrimination in patients aged ≥60 years, with an AUC of 0.760 (95% CI: 0.692–0.828). Although the sensitivity remained high (0.855), the specificity decreased to 0.491 in this subgroup.

Collectively, these findings demonstrate that integrating DLR features with clinical variables improves CHD risk prediction in MAFLD patients, providing a foundation for further interpretation of the clinical implications and potential applications of the proposed model.

4. Discussion

To our knowledge, this study represents one of the first attempts to develop a deep learning radiomics nomogram based on AI-assisted segmentation for CHD risk stratification in patients with MAFLD. Previous studies have developed clinical prediction models and nomograms to estimate CHD risk in MAFLD populations (8), but imaging-based deep learning radiomics approaches remain insufficiently explored. Recent studies have demonstrated that integrating deep learning features, radiomics signatures, and clinical variables can significantly improve predictive performance compared with single-modality models (11, 20–22). Therefore, by integrating radiomics features and DL features extracted from non-contrast liver CT with clinical risk factors, we established a combined DLRC model that exhibited excellent discrimination, calibration, and clinical utility in predicting the risk of CHD in patients with MAFLD. The combined model achieved AUCs of 0.917 and 0.878 in the training and test cohorts, respectively, outperforming the clinical, radiomics, and DLR models individually. These findings indicate that quantitative liver imaging features combined with clinical characteristics may offer valuable information for predicting CHD risk in MAFLD patients.

MAFLD is increasingly recognized as a multisystem metabolic disorder rather than a liver-confined disease (1, 23). Accumulating evidence shows that cardiovascular disease is the leading cause of mortality in MAFLD patients (5, 24, 25). In our study, multivariable logistic regression identified advanced age, hypertension, diabetes, hyperlipidemia, and male sex as independent predictors of CHD in MAFLD patients, consistent with previous epidemiological studies (26–29), indicating that traditional cardiometabolic risk factors play key roles in coronary atherosclerosis development. Among these, hypertension had the strongest predictive effect (OR = 2.541), underscoring the importance of blood pressure control. The associations of diabetes and hyperlipidemia with CHD may reflect shared pathophysiological mechanisms, including IR, dyslipidemia, chronic inflammation, oxidative stress, and endothelial dysfunction (4, 30–32). Interestingly, lower BMI was associated with increased CHD risk, which may be related to age-related muscle loss, body composition changes, or the “obesity paradox” (33–36); however, this finding requires further investigation and should be interpreted cautiously.

Radiomics has emerged as a promising tool for extracting high-dimensional quantitative data from medical images. In our study, the radiomics model achieved AUCs of 0.821 and 0.784 in the training and test cohorts, suggesting that liver CT images harbor imaging features associated with CHD risk. Prior studies have confirmed that hepatic steatosis is closely linked to systemic metabolic dysfunction and cardiovascular disease (37–39). Radiomics analysis can detect subtle changes in liver morphology, textural heterogeneity, and tissue microstructure imperceptible to the human eye (40, 41). These imaging features may indirectly reflect the metabolic abnormalities and inflammatory processes underlying coronary heart disease development.

Compared with conventional radiomics, DL has the advantage of automatically learning complex hierarchical representations from imaging data without relying on handcrafted feature engineering (42). In our study, the DLR model, which integrated radiomics features and compressed DL features extracted using DenseNet-121, achieved significantly superior predictive performance compared with the radiomics model alone. The DLR model achieved AUCs of 0.895 and 0.854 in the training and test cohorts, respectively. This improvement suggests that DL features provide complementary information beyond traditional radiomics descriptors. While radiomics primarily captures shape, intensity, and texture features from CT images, DL can identify highly abstract imaging patterns and nonlinear relationships that may be associated with CHD risk (43). Similar findings have been reported in previous studies involving lung cancer, gastric cancer, and rectal cancer, where the fusion of radiomics and DL features consistently outperformed either approach alone (44–46).

In our nomogram, the DLR score contributed substantially to the total points, suggesting that imaging-derived deep features offer significant incremental predictive value beyond traditional clinical risk factors. This indicates that the DLR score may capture high-dimensional patterns and abstract phenotypes related to metabolic dysfunction, hepatic microstructural remodeling, and systemic atherosclerotic burden that are not adequately reflected in routine clinical variables. By integrating complementary information from clinical and imaging domains, the proposed nomogram provides a more comprehensive characterization of disease prediction. Moreover, the DeLong test confirmed that the combined model significantly outperformed the other models, supporting the incremental value of clinical and imaging data fusion. In the subgroup analysis, the combined model demonstrated consistent performance across sex subgroups but showed reduced specificity in patients aged ≥60 years. This finding may be partially explained by the application of a uniform classification threshold across different age groups. Since older individuals often have a higher burden of cardiovascular risk factors (47), a fixed threshold may affect model calibration and increase false-positive predictions in this subgroup. Future studies will be further explored with larger cohorts and external validation datasets.

Another notable strength of this study is the adoption of an AI-assisted segmentation strategy. Manual liver segmentation is time-consuming and subject to inter-observer variability, which limits the practical application of radiomics in clinical settings. By employing a deep learning-assisted segmentation framework integrated within ITK-SNAP, we achieved rapid and accurate liver delineation with minimal manual adjustment. This workflow substantially reduced labor costs and improved reproducibility, thereby enhancing the feasibility of large-scale clinical implementation. Recent studies have demonstrated that automatic segmentation facilitates efficient radiomics analysis while maintaining high predictive accuracy (48, 49).

From a clinical perspective, the proposed nomogram represents a non-invasive and easily accessible tool for assessing CHD risk in patients with MAFLD. Given that liver CT examinations are routinely performed in many clinical centers, our fusion model can generate individualized risk estimates for CHD without the need for additional imaging modalities or invasive diagnostic procedures. The satisfactory calibration performance and DCA suggest that the model provides a meaningful net clinical benefit across a broad range of threshold probabilities. Accordingly, this approach holds potential for identifying high-risk individuals who may benefit from early preventive interventions.

5. Limitations

Several limitations of this study should be acknowledged. First, this was a retrospective single-center study, which may have introduced selection bias and limited the generalizability of the findings. External multicenter validation is necessary before clinical implementation. Second, although AI-assisted liver segmentation using nnInteractive improved segmentation efficiency and reduced the workload of manual delineation, visual inspection and occasional minor manual corrections were still required in some cases. Further optimization and validation of fully automated segmentation workflows are warranted in future studies. Third, only non-contrast liver CT images were analyzed; incorporating other imaging modalities such as MRI or ultrasound may further improve predictive performance.

6. Conclusion

We developed and validated an AI-assisted segmentation-based deep learning radiomics nomogram for identifying individuals at risk of CHD among patients with MAFLD. By integrating clinical parameters, radiomics features, and deep learning features, the proposed DLRC model demonstrated superior predictive performance compared with individual models. This non-invasive and efficient approach may facilitate early CHD risk stratification and support personalized cardiovascular risk management in patients with MAFLD. Future studies involving multicenter external validation, larger prospective cohorts, and integration into clinical decision-support systems are warranted to further evaluate the generalizability, robustness, and clinical utility of the proposed model.

Funding Statement

The author(s) declared that funding support was not received for this work and/or its publication.

Footnotes

Edited by: Syed Immamul Ansarullah, University of Kashmir, India

Reviewed by: Mudita Uppal, IILM University, India

Mudasir Kirmani, Sher-e-Kashmir University of Agricultural Sciences and Technology of Kashmir, India

Md Aminul Islam, Ulster University, United Kingdom

Data availability statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Ethics statement

The studies involving humans were approved by the Ethics Committee of First Affiliated Hospital of Huzhou Normal University (Approval Number: 2025KYLL099-01). The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants’ legal guardians/next of kin because Prior to data analysis, all datasets underwent thorough anonymization, with direct personal identifiers being irreversibly removed. Given the retrospective design of the study, the ethics committee granted a formal waiver of informed consent.

Author contributions

LQ: Conceptualization, Data curation, Formal analysis, Writing – original draft. SZ: Methodology, Validation, Writing – review & editing. BW: Software, Validation, Writing – original draft. MY: Investigation, Resources, Writing – original draft. YL: Data curation, Writing – original draft. BL: Formal analysis, Writing – original draft. JX: Investigation, Writing – original draft. XL: Conceptualization, Project administration, Supervision, Writing – review & editing.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fendo.2026.1908048/full#supplementary-material

Supplementaryfile1.docx (15.7KB, docx)

References

  • 1. Eslam M, Newsome PN, Sarin SK, Anstee QM, Targher G, Romero-Gomez M, et al. A new definition for metabolic dysfunction-associated fatty liver disease: An international expert consensus statement. J Hepatol. (2020) 73:202–9. doi:  10.1016/j.jhep.2020.03.039 [DOI] [PubMed] [Google Scholar]
  • 2. Wild SH, Lazarus JV, Spearman CW, Ocama P, Bilson J, Zhou XD, et al. MASLD prevalence, incidence and global aspects. Diabetologia. (2026). doi:  10.1007/s00125-026-06726-1 [DOI] [PubMed] [Google Scholar]
  • 3. Xu B, Li H, Pi Y, Li K. Global, regional and national burden of metabolic dysfunction-associated steatotic liver disease in adolescents and adults aged 15-49 years from 1990 to 2021: results from the 2021 Global Burden of Disease study. Front Med. (2025) 12:1568211. doi:  10.3389/fmed.2025.1568211 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 4. Badmus OO, Hinds TD, Jr., Stec DE. Mechanisms linking metabolic-associated fatty liver disease (MAFLD) to cardiovascular disease. Curr Hypertension Rep. (2023) 25:151–62. doi:  10.1007/s11906-023-01242-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 5. Yang J, Kim YR, Na SK, Kim S, An J, Shim JH. Mortality and cardiovascular outcomes in patients with MAFLD compared with patients with MASLD: A systematic review and meta-analysis. Gut Liver. (2026) 20:137–52. doi:  10.5009/gnl250275 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 6. Moon JH, Jeong S, Jang H, Koo BK, Kim W. Metabolic dysfunction-associated steatotic liver disease increases the risk of incident cardiovascular disease: a nationwide cohort study. EClinicalMedicine. (2023) 65:102292. doi:  10.1016/j.eclinm.2023.102292 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7. Lee HH, Lee HA, Kim EJ, Kim HY, Kim HC, Ahn SH, et al. Metabolic dysfunction-associated steatotic liver disease and risk of cardiovascular disease. Gut. (2024) 73:533–40. doi:  10.1136/gutjnl-2023-331003 [DOI] [PubMed] [Google Scholar]
  • 8. Li Z, Chen X, Wang J, Chen W, Zhang R, Cao L, et al. Development of a nomogram model for predicting coronary heart disease in patients with metabolic-associated fatty liver disease. Front Cardiovasc Med. (2025) 12:1652321. doi:  10.3389/fcvm.2025.1652321 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 9. Kim M, Lee HA, Kim J, Kang HT. Metabolic dysfunction-associated steatotic liver disease and incident coronary events with revascularization: A nationwide cohort study. Atherosclerosis. (2026) 415:120698. doi:  10.1016/j.atherosclerosis.2026.120698 [DOI] [PubMed] [Google Scholar]
  • 10. Zhao J, Fan H, Wang T, Yu B, Mao S, Wang X, et al. TyG index is positively associated with risk of CHD and coronary atherosclerosis severity among NAFLD patients. Cardiovasc Diabetol. (2022) 21:123. doi:  10.1186/s12933-022-01548-y [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11. Ding N, Jin L, Yin S, Ji Y, Wang X, Li M. A multicenter study of automatic segmentation-based multimodal fusion integrating radiomics, deep learning, and clinical parameters for prostate cancer detection. Abdominal Radiol (New York). (2026). doi:  10.1007/s00261-026-05548-4 [DOI] [PubMed] [Google Scholar]
  • 12. Yan L, Xu J, Ye X, Lin M, Gong Y, Fang Y, et al. Development and validation of ultrasound-based radiomics deep learning model to identify bone erosion in rheumatoid arthritis. Clin Rheumatol. (2025) 44:2635–45. doi:  10.1007/s10067-025-07481-1 [DOI] [PubMed] [Google Scholar]
  • 13. Mayerhoefer ME, Materka A, Langs G, Häggström I, Szczypiński P, Gibbs P, et al. Introduction to radiomics. J Nucl Med Off Publication Soc Nucl Med. (2020) 61:488–95. doi:  10.2967/jnumed.118.222893 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14. Kanakarajan H, De Baene W, Hanssens P, Sitskoorn M. Predicting local control of brain metastases after stereotactic radiotherapy with clinical, radiomics and deep learning features. Radiat Oncol (London England). (2024) 19:182. doi:  10.1186/s13014-024-02573-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15. Yin X, Lu Y, Cui Y, Zhou Z, Wen J, Huang Z, et al. CT-based radiomics-deep learning model predicts occult lymph node metastasis in early-stage lung adenocarcinoma patients: A multicenter study. Chin J Cancer Res = Chung-Kuo Yen Cheng Yen Chiu. (2025) 37:12–27. doi:  10.21147/j.issn.1000-9604.2025.01.02 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16. Li Z, Qin Y, Liao X, Wang E, Cai R, Pan Y, et al. Comparison of clinical, radiomics, deep learning, and fusion models for predicting early recurrence in locally advanced rectal cancer based on multiparametric MRI: a multicenter study. Eur J Radiol. (2025) 189:112173. doi:  10.1016/j.ejrad.2025.112173 [DOI] [PubMed] [Google Scholar]
  • 17. Guo Y, Yin H, Zhang H, Liang P, Gao J, Cheng M. Combining radiomics and deep learning to predict liver metastasis of gastric cancer on CT image. Front Oncol. (2025) 15:1613972. doi:  10.3389/fonc.2025.1613972 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18. Zhao X, Wang M, Wei Y, Lu Z, Peng Y, Cheng X, et al. Overview of multimodal radiomics and deep learning in the prediction of axillary lymph node status in breast cancer. Acad Radiol. (2025) 32:6623–41. doi:  10.1016/j.acra.2025.07.017 [DOI] [PubMed] [Google Scholar]
  • 19. Vascular Disease Branch of China Medical Biotechnology Association. Chinese Research Hospital Association of Hepatology (Integrated Traditional Chinese and Western Medicine) Professional Committee. Chinese Society of General Practice of Chinese Medical Association . The primary care guidelines for the diagnosis and management of metabolic dysfunction-associated fatty liver disease (2025). Chin J Hepatol. (2025) 33:422–33. Available online at: https://kns.cnki.net/kcms2/article/abstract?v=7jvqSXIa2LUxwWB_iHUVweFRWmFb5-Q_nAV16GtNKQ94n9nig4pUoB_YJ8IxiaLJ-95xfVSdYIRhSYy9gM5xDSDrYL8qdKBSNM19c1VO8v7QyeOp06xMegeyJuwzo6NYbFd9q0DXPhEj5cQNAKRBdSsnAA5MIUWCyTlt1j0RE3QhgZcMyaJtAg==&uniplatform=NZKPT&language=CHS (Accessed July 10, 2026). [Google Scholar]
  • 20. Zhao J, Wang T, Wang B, Satishkumar BM, Ding L, Sun X, et al. Deep learning radiomics fusion model to predict visceral pleural invasion of clinical stage IA lung adenocarcinoma: a multicenter study. J Cardiothoracic Surg. (2025) 20:246. doi:  10.1186/s13019-025-03488-6 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 21. Cai W, Wu X, Guo K, Chen Y, Shi Y, Lin X. Deep-learning, radiomics and clinic based fusion models for predicting response to infliximab in Crohn's disease patients: A multicentre, retrospective study. J Inflammation Res. (2024) 17:7639–51. doi:  10.2147/jir.S484485 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22. Zhang H, Yang YF, Song XL, Hu HJ, Yang YY, Zhu X, et al. An interpretable artificial intelligence model based on CT for prognosis of intracerebral hemorrhage: a multicenter study. BMC Med Imaging. (2024) 24:170. doi:  10.1186/s12880-024-01352-y [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23. Byrne CD, Targher G. NAFLD: a multisystem disease. J Hepatol. (2015) 62:S47–64. doi:  10.1016/j.jhep.2014.12.012 [DOI] [PubMed] [Google Scholar]
  • 24. Targher G, Byrne CD, Tilg H. NAFLD and increased risk of cardiovascular disease: clinical associations, pathophysiological mechanisms and pharmacological implications. Gut. (2020) 69:1691–705. doi:  10.1136/gutjnl-2020-320622 [DOI] [PubMed] [Google Scholar]
  • 25. Lin H, Zhang X, Li G, Wong GL, Wong VW. Epidemiology and clinical outcomes of metabolic (dysfunction)-associated fatty liver disease. J Clin Trans Hepatol. (2021) 9:972–82. doi:  10.14218/jcth.2021.00201 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 26. Akeno Y, Maeda M, Murata F, Fukuda H. Metabolic dysfunction-associated fatty liver disease and risks of incident coronary artery disease and cerebrovascular disease: LIFE study. Hepatol Res Off J Japan Soc Hepatol. (2025) 55:807–17. doi:  10.1111/hepr.14185 [DOI] [PubMed] [Google Scholar]
  • 27. Zhang Y, Xia Z, Cai X, Su X, Jin A, Mei L, et al. Association of metabolic dysfunction-associated fatty liver disease with systemic atherosclerosis: a community-based cross-sectional study. Cardiovasc Diabetol. (2023) 22:342. doi:  10.1186/s12933-023-02083-0 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 28. Nishihara T, Miyoshi T, Nakashima M, Miki T, Toda H, Yoshida M, et al. Prognostic value of metabolic dysfunction-associated steatotic liver disease over coronary computed tomography angiography findings: comparison with no-alcoholic fatty liver disease. Cardiovasc Diabetol. (2024) 23:167. doi:  10.1186/s12933-024-02268-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 29. Drożdż K, Nabrdalik K, Kwiendacz H, Hendel M, Olejarz A, Tomasik A, et al. Risk factors for cardiovascular disease in patients with metabolic-associated fatty liver disease: a machine learning approach. Cardiovasc Diabetol. (2022) 21:240. doi:  10.1186/s12933-022-01672-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30. Mostafa AM, Pan Z, Yu ML, Örmeci N, Fouad Y, Eslam M. MAFLD: a comprehensive review of the link between metabolic dysfunction and cardiovascular risk. Hepatic Med Evidence Res. (2025) 17:75–90. doi:  10.2147/hmer.S506402 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31. Fan Y, Yan Z, Li T, Li A, Fan X, Qi Z, et al. Primordial drivers of diabetes heart disease: comprehensive insights into insulin resistance. Diabetes Metab J. (2024) 48:19–36. doi:  10.4093/dmj.2023.0110 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32. Yang DR, Wang MY, Zhang CL, Wang Y. Endothelial dysfunction in vascular complications of diabetes: a comprehensive review of mechanisms and implications. Front Endocrinol. (2024) 15:1359255. doi:  10.3389/fendo.2024.1359255 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 33. Liu C, Wong PY, Chung YL, Chow SK, Cheung WH, Law SW, et al. Deciphering the "obesity paradox" in the elderly: a systematic review and meta-analysis of sarcopenic obesity. Obes Rev Off J Int Assoc For Study Obes. (2023) 24:e13534. doi:  10.1111/obr.13534 [DOI] [PubMed] [Google Scholar]
  • 34. Jiang M, Ren X, Han L, Zheng X. Associations between sarcopenic obesity and risk of cardiovascular disease: a population-based cohort study among middle-aged and older adults using the CHARLS. Clin Nutr (Edinburgh Scotland). (2024) 43:796–802. doi:  10.1016/j.clnu.2024.02.002 [DOI] [PubMed] [Google Scholar]
  • 35. Eitmann S, Matrai P, Hegyi P, Balasko M, Eross B, Dorogi K, et al. Obesity paradox in older sarcopenic adults - a delay in aging: a systematic review and meta-analysis. Ageing Res Rev. (2024) 93:102164. doi:  10.1016/j.arr.2023.102164 [DOI] [PubMed] [Google Scholar]
  • 36. Prado CM, Batsis JA, Donini LM, Gonzalez MC, Siervo M. Sarcopenic obesity in older adults: a clinical overview. Nat Rev Endocrinol. (2024) 20:261–77. doi:  10.1038/s41574-023-00943-z [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 37. Yanai H, Adachi H, Hakoshima M, Iida S, Katsuyama H. Metabolic-dysfunction-associated steatotic liver disease-its pathophysiology, association with atherosclerosis and cardiovascular disease, and treatments. Int J Mol Sci. (2023) 24:15473. doi:  10.3390/ijms242015473 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 38. Zheng H, Sechi LA, Navarese EP, Casu G, Vidili G. Metabolic dysfunction-associated steatotic liver disease and cardiovascular risk: a comprehensive review. Cardiovasc Diabetol. (2024) 23:346. doi:  10.1186/s12933-024-02434-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39. Vaz K, Kemp W, Majeed A, Lubel J, Magliano DJ, Glenister KM, et al. NAFLD and MAFLD independently increase the risk of major adverse cardiovascular events (MACE): a 20-year longitudinal follow-up study from regional Australia. Hepatol Int. (2024) 18:1135–43. doi:  10.1007/s12072-024-10706-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40. Maino C, Vernuccio F, Cannella R, Franco PN, Giannini V, Dezio M, et al. Radiomics and liver: Where we are and where we are headed? Eur J Radiol. (2024) 171:111297. doi:  10.1016/j.ejrad.2024.111297 [DOI] [PubMed] [Google Scholar]
  • 41. Wang XM, Zhang XJ. Role of radiomics in staging liver fibrosis: a meta-analysis. BMC Med Imaging. (2024) 24:87. doi:  10.1186/s12880-024-01272-x [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 42. Rundo L, Militello C. Image biomarkers and explainable AI: handcrafted features versus deep learned features. Eur Radio Exp. (2024) 8:130. doi:  10.1186/s41747-024-00529-y [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 43. Cè M, Chiriac MD, Cozzi A, Macrì L, Rabaiotti FL, Irmici G, et al. Decoding radiomics: a step-by-step guide to machine learning workflow in hand-crafted and deep learning radiomics studies. Diagnostics (Basel Switzerland). (2024) 14:2473. doi:  10.3390/diagnostics14222473 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44. Yang Y, Han K, Xu Z, Cai Z, Zhao H, Hong J, et al. Development and validation of multiparametric MRI-based interpretable deep learning radiomics fusion model for predicting lymph node metastasis and prognosis in rectal cancer: a two-center study. Acad Radiol. (2025) 32:2642–54. doi:  10.1016/j.acra.2024.11.045 [DOI] [PubMed] [Google Scholar]
  • 45. Jiang Z, Xie W, Zhou X, Pan W, Jiang S, Zhang X, et al. A virtual biopsy study of microsatellite instability in gastric cancer based on deep learning radiomics. Insights into Imaging. (2023) 14:104. doi:  10.1186/s13244-023-01438-1 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 46. Zhang X, Zhang G, Qiu X, Yin J, Tan W, Yin X, et al. Exploring non-invasive precision treatment in non-small cell lung cancer patients through deep learning radiomics across imaging features and molecular phenotypes. biomark Res. (2024) 12:12. doi:  10.1186/s40364-024-00561-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47. Lind L, Sundström J, Ärnlöv J, Lampa E. Impact of aging on the strength of cardiovascular risk factors: a longitudinal study over 40 years. J Am Heart Assoc. (2018) 7:e007061. doi:  10.1161/jaha.117.007061 [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48. Stefano A, Bini F, Lauciello N, Pasini G, Marinozzi F, Russo G. Implementation of automatic segmentation framework as preprocessing step for radiomics analysis of lung anatomical districts. BioMedInformatics. (2024) 4:2309–20. doi:  10.3390/biomedinformatics4040125 30654563 [DOI] [Google Scholar]
  • 49. Xiong L, Tang X, Jiang X, Chen H, Qian B, Chen B, et al. Automatic segmentation-based multi-modal radiomics analysis of US and MRI for predicting disease-free survival of breast cancer: a multicenter study. Breast Cancer Res BCR. (2024) 26:157. doi:  10.1186/s13058-024-01909-3 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementaryfile1.docx (15.7KB, docx)

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Articles from Frontiers in Endocrinology are provided here courtesy of Frontiers Media SA

RESOURCES