A fusion model of deep learning and conventional features based on computed tomography angiography of carotid plaque for predicting the risk of acute ischemic stroke

Introduction

Stroke remains one of the leading causes of global mortality and long-term disability, with ischemic stroke constituting the majority of the reported cases (1). Carotid atherosclerosis represents a major underlying etiology of stroke, responsible for approximately 15–20% of all ischemic events (2). Historically, stroke risk stratification has been predominantly based on the degree of luminal stenosis. However, a growing body of evidence indicates that specific vulnerable plaque characteristics, such as intraplaque hemorrhage (IPH), large lipid-rich necrotic core (LRNC), thin fibrous cap, and inflammatory infiltration, are more strongly associated with cerebrovascular events than is stenosis severity alone (3,4). Therefore, the accurate identification of high-risk plaques, particularly those implicated in acute stroke, is crucial for advancing preventive strategies.

Computed tomography angiography (CTA), magnetic resonance imaging (MRI), and ultrasound are the most widely used noninvasive imaging modalities for the evaluation of vulnerable plaques and related clinical decision-making. Among these, CTA is frequently recommended as the initial imaging tool due to its rapid acquisition, widespread availability, and its ability to provide valuable diagnostic information. A few studies (5-7) have preliminarily investigated the use of CTA for examining plaque. Although CTA demonstrates high accuracy in detecting high-risk plaque features such as IPH, LRNC, and ulcerations, it has been reported that imaging-derived compositional characteristics, including IPH, are not consistently associated with acute stroke risk (8,9). Therefore, the relationship between imaging plaque features and clinical outcomes remains incompletely understood.

Deep learning methods have emerged as powerful tools for automatically extracting hierarchical feature representations directly from medical images. These data-driven features can be effectively leveraged for tasks such as disease identification and risk stratification (10-12). For instance, convolutional neural networks (CNNs) applied to ultrasound imaging have shown promise in discriminating vulnerable from stable carotid plaques (13). However, CNNs inherently tend to capture global patterns and may lack sensitivity to localized pathological findings such as focal vulnerable plaques or heterogeneous tumors that are confined to specific anatomical regions. In contrast, Swin Transformers represent a promising alternative, as they incorporate a window-based self-attention mechanism and a shifted window strategy (14). They thus facilitate the efficient integration of global contextual information and localized detail, a capability particularly beneficial for the analysis of high-resolution medical images.

Building on these technical insights, this study aimed to develop and validate a multimodal fusion model that integrates clinical risk factors with Swin Transformer-derived imaging features extracted from carotid plaque CTA for predicting ipsilateral acute ischemic stroke. Furthermore, Shapley additive explanations (SHAP)-based interpretability methods were employed to identify the most influential clinical and deep learning features, with the goal of establishing a transparent and clinically actionable framework for risk management in stroke. We present this article in accordance with the TRIPOD + AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2688/rc).

Methods

Patients

This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments and was approved by the Ethics Committee of The First Affiliated Hospital of Chongqing Medical University (No. K2023-388). The requirement for informed consent was waived due to the retrospective nature of the analysis.

We screened patients who underwent head and neck CTA and diffusion-weighted imaging (DWI) at The First Affiliated Hospital of Chongqing Medical University between January 2020 and April 2025. The inclusion criteria were as follows: (I) presence of measurable plaque in the carotid bifurcation or extracranial internal carotid artery (ICA) as confirmed by CTA; (II) for patients with multiple plaques, vulnerable plaque located in the carotid bifurcation or extracranial ICA according to the Carotid Plaque Reporting and Data System (Plaque-RADS) criteria (15); and (III) a time interval between head and neck CTA and brain DWI examinations of no more than 7 days. Meanwhile, the exclusion criteria were as follows: (I) insufficient clinical data; (II) history of cervical radiotherapy, cerebral hemorrhage, tumor, trauma, or prior brain surgery; (III) posterior circulation stroke, isolated lacunar infarction, or severe intracranial arterial stenosis; (IV) suspected cardioembolic source; (V) nonatherosclerotic intracranial arterial disease such as vasculitis, moyamoya disease, dissection, or reversible cerebral vasoconstriction syndrome; (VI) prior carotid stenting or endarterectomy; and (VII) poor image quality.

Patients with cerebral ischemic symptoms and DWI high signal on head MRI within 2 weeks of symptom onset were placed into the symptom group. Meanwhile, patients with no cerebral ischemic symptoms and no DWI high signal on head MRI were placed into the asymptomatic group. We excluded patients with typical transient ischemic attack.

A total of 264 patients were ultimately enrolled and categorized into symptomatic (positive DWI with ipsilateral internal carotid plaque) or asymptomatic (negative DWI with measurable plaque) groups. Eligible patients were randomly allocated to a training cohort (n=184) and test cohort (n=80) at a 7:3 ratio. The allocation of the two groups can be found in Table S1. The patient selection process is summarized in Figure 1.

Figure 1 Flowchart of patient recruitment. CTA, computed tomography angiography; MRI, magnetic resonance imaging.

Collection of clinical information

Demographic, clinical, and laboratory data were collected for each patient, including age, gender, body mass index (BMI), smoking status, alcohol consumption, hypertension, diabetes, hyperlipidemia, total cholesterol, triglycerides, high-density lipoprotein cholesterol (HDL-C), and low-density lipoprotein cholesterol (LDL-C).

Imaging acquisition

All patients underwent head and neck CTA at a single center, with four scanners being used: the Aquilion ONE (Canon Medical Systems, Otawara, Japan), the Discovery CT750 HD (GE HealthCare, Chicago, IL, USA), the SOMATOM Definition AS+ (Siemens Healthineers, Erlangen, Germany), and the SOMATOM Force (Siemens Healthineers). The patient was placed in the supine position. Iodinated contrast medium (iopamidol; Bracco Sine Pharmaceutica, Shanghai, China) was administered intravenously via the antecubital vein at a dose of 45–65 mL (based on body weight: 350–400 mL/kg) with an infusion rate of 4–5 mL/s, which was followed by a 50-mL saline flush. The scan range extended from the aortic arch to the cranial vault in a foot-to-head direction. The scan parameters can be found in Table S2.

Clinical imaging analysis

Two radiologists with 3 years of neuroimaging experience and blinded to the clinical information reconstructed carotid arteries using three-dimensional (3D) Slicer software and assessed the following conventional plaque characteristics: (I) luminal stenosis as measured via the North American Symptomatic Carotid Endarterectomy Trial (NASCET) diameter stenosis percentage method, i.e., [1− (narrowest diameter/normal diameter)] ×100% (16); (II) maximum plaque thickness in axial CTA images (17); (III) plaque ulceration, defined as contrast extension into the plaque ≥1 mm in any plane; (IV) carotid rim sign, defined as adventitial calcification (thickness <2 mm) with internal soft plaque (>2 mm thick) (18); (V) remodeling index, calculated as (vessel area at maximum stenosis/distal reference vessel area) ×100% (19); (VI) plaque burden, calculated as (plaque area/vessel area) ×100%; (VII) carotid artery plaque risk score (Plaque-RADS) (15), with scores ≥3 classified as high risk and <3 as low risk.

Region-of-interest (ROI) segmentation

A standardized image processing pipeline was applied to all images. ROI segmentation was performed with 3D Slicer software version 5.6.2. ROIs were delineated at the arterial phase at the level of maximum plaque area (19,20). Two junior radiologists with 3 years of neuroradiology experience independently performed the segmentations, which were subsequently reviewed by a senior radiologist with 25 years of experience. Discrepancies were resolved through consensus discussion. For plaques involving both internal and external carotid arteries, the ICA segment was prioritized. Representative imaging features of a symptomatic carotid plaque are shown in Figure 2.

Figure 2 Case presentation of a patient with acute ischemic stroke. A 69-year-old female presented with left upper-limb weakness for 5 days and left lower-limb weakness for 2 days. (A,B) Axial CTA of the carotid artery revealed a thick plaque at the proximal segment of the right internal carotid artery with a positive rim sign (long arrow). (C,D) Axial diffusion-weighted imaging showed acute infarction areas. (E) Sagittal reconstructed CTA of the carotid artery showed carotid lumen stenosis and plaque distribution (long arrow). CTA, computed tomography angiography.

Only one plaque from each patient was ultimately selected for subsequent analysis. Specifically, symptomatic plaques were defined as the plaques in the relevant vascular area of symptomatic patients, and asymptomatic plaques were considered those in asymptomatic patients. Subsequently, carotid plaques were scored according to the Plaque-RADS criteria (15,21). If multiple plaques were present on the symptomatic side, the plaque with the highest-risk features was selected based on Plaque-RADS criteria. For asymptomatic patients with bilateral carotid plaques, the side with a higher Plaque-RADS score was designated as the target vessel; in cases of identical scores on both sides, the side with a thicker plaque was chosen. For asymptomatic patients with unilateral carotid plaques, the affected side was directly selected as the study side.

Feature extraction, feature selection, and model construction

Deep learning feature extraction

The maximum plaque area identified on CTA imaging was resized to a 224×224 resolution via bilinear interpolation and was used as the input image for the deep learning models. The primary architecture consisted of a pretrained Swin Tiny Transformer enhanced with attention pooling layers, along with a classification head incorporating progressive dropout (0.3 to 0.1); a multilayer perceptron block integrated with layer normalization and Gaussian error linear unit activation was also embedded in the architecture to optimize feature transformation and propagation. Feature projection modules maintained consistent 1,024-dimensional output representations throughout the network. For comparative analysis, we additionally implemented a DenseNet121 architecture following identical preprocessing protocols. Both models were trained with focal loss (α=1 and γ=2) with AdamW optimization (initial learning rate=1e−4; weight decay=1e−4) over 200 epochs. Training incorporated reduced learning rate on plateau scheduling coupled with early stopping (patience =20) to prevent overfitting. The automated processing pipeline generated standardized 1,024-dimensional feature vectors while producing multipanel visualizations highlighting discriminative image regions through gradient-weighted class activation mapping (Grad-CAM), enabling both quantitative analysis and qualitative assessment of decision patterns.

Feature processing and predictive model development were conducted according to a structured pipeline. Initially, all conventional and deep learning features underwent z-score normalization. Feature selection was then performed through sequential application of Pearson correlation analysis, statistical testing (t-test or Mann-Whitney test based on distribution characteristics), and least absolute shrinkage and selection operator (LASSO) regression. Using the refined feature sets, we constructed three distinct model categories: conventional models (utilizing only clinical risk factors and standard imaging features), deep learning models (utilizing only deep learning-derived features), and fusion models (integrating both feature types).

Model evaluation and interpretation

To select the best performance of the model for our classification task, we employed four machine learning classifiers, including support vector machine (SVM), random forest, stochastic gradient descent, and k-nearest neighbors. The classifier that demonstrated the best performance for each feature set was subsequently selected to construct the conventional model, DenseNet121 model, Swin Transformer model, and Fusion model, respectively.

Model performance was comprehensively assessed according to the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Decision curve analysis (DCA) was performed to assess the clinical net benefit of the models. To enhance clinical interpretability, we applied SHAP analysis, which quantifies feature contributions to individual predictions based on game theory, thereby identifying the most influential predictors and providing global feature importance rankings. The workflow of the model development is shown in Figure 3.

Figure 3 Flowchart of model construction. AUC, area under the receiver operating characteristic curve; DCA, decision curve analysis; Grad-CAM, gradient-weighted class activation mapping; ROI, region-of-interest.

Statistical analysis

Statistical analysis was performed with R software (The R Foundation for Statistical Computing, Vienna, Austria). Categorical variables are presented as frequencies (%), while continuous variables are expressed as the mean ± standard deviation or as the median and interquartile range. Categorical variables were compared with the χ² test, while continuous variables were compared with the Student t-test or Mann-Whitney test. Additionally, the DeLong test was employed to compare the AUC values of the different models, with a P value <0.05 being considered statistically significant.

Results

Baseline characteristics

After simple random sampling, the training set comprised 100 asymptomatic patients and 84 symptomatic patients, while the test set comprised 45 asymptomatic patients and 35 symptomatic patients. The clinical information and CTA imaging characteristics of the patients are summarized in Table 1. Patients in the symptomatic group exhibited significantly higher risk levels for clinical factors—including smoking, drinking, diabetes, and LDL-C level, and for imaging characteristics—including degrees of stenosis, plaque thickness, plaque burden, ulceration, rim sign, and high-risk Plaque-RADS score.

Key features

This study assessed two deep learning models, DenseNet121 and Swin Transformer. Training set images were input into the models for feature extraction, and multipanel visualizations were generated via Grad-CAM (Figure 4). Ultimately, DenseNet121 and Swin Transformer extracted 1,024 features each. Following variable selection, LASSO regression was applied to both conventional and deep learning features. The key conventional features retained were diabetes, degree of stenosis, plaque burden, ulceration, rim sign, plaque thickness, and Plaque-RADS score, with 15 selected features from DenseNet121 and 16 selected features from Swin Transformer. The weight distributions of all selected features are visualized in Figure 5.

Figure 4 Grad-CAM visualization for model interpretation. (A) The activation map was generated by a symptomatic carotid plaque. (B) The activation map generated by the asymptomatic carotid plaque. Grad-CAM activation maps are predominantly highlighted in the low-density regions of the plaque. Grad-CAM, gradient-weighted class activation mapping.

Figure 5 Key features of conventional and deep learning after LASSO selection. (A) Conventional features. (B) DenseNet121 features. (C) Swin Transformer features. DM, diabetes mellitus; LASSO, least absolute shrinkage and selection operator; PB, plaque burden; RADS, plaque reporting and data system score.

Development and performance of models

Among the four classifiers, SVM demonstrated the highest accuracy and AUC (Table S3). Based on the selected feature set described above, we constructed conventional, DenseNet121, and Swin Transformer models using SVM and fused the conventional and Swin Transformer features to build a fusion model. In the training and validation cohorts, the AUCs were, respectively, 0.729 and 0.691 for the conventional model, 0.975 and 0.954 for the DenseNet121 model, 0.956 and 0.997 for the Swin Transformer model, and 0.980 and 0.989 for the fusion model (Figure 6 and Table 2). The DeLong test results indicated that the deep learning models and fusion model significantly outperformed the conventional model in both the training and validation cohorts (P<0.05). No significant difference was observed between the deep learning models and fusion model in either the training or validation cohorts. Notably, Swin Transformer exhibited better performance in the validation cohort than in the training cohort and had an excellent classification performance in the validation cohort. DCA further revealed its favorable clinical utility. Additionally, the calibration curve indicated that the predicted probabilities of the Swin Transformer model were well calibrated, suggesting the high reliability of the model (Figure 6).

Figure 6 Performance evaluation of the four models in the training and validation cohorts. (A,C,E) The ROC curves, decision curves, and calibration curves for the training cohort, respectively. (B,D,F) The ROC curves, decision curves, and calibration curves for the validation cohort, respectively. AUC, area under the receiver operating characteristic curve; ROC, receiver operating characteristic.

Discussion

This study developed and validated a multimodal fusion model that integrates clinical imaging characteristics and the Swin Transformer features derived from two-dimensional (2D) head and neck CTA imaging to predict carotid plaque associated with acute ischemic stroke. Both the fusion model and Swin Transformer model demonstrated excellent performance on an independent validation cohort. Their AUC values significantly outperformed those of the model relying solely on clinical imaging features and thus represent a robust tool for risk stratification in patients with carotid plaques.

The conventional model in this study yielded an AUC of 0.691 in the validation cohort, a performance that is within the range reported in related work (7), yet it also highlights the limitations of relying exclusively on traditional risk factors and conventional imaging parameters for risk prediction. A number of previous studies have confirmed that traditional risk factors, such as diabetes, are significantly associated with symptomatic carotid plaques (22-25). Our study also confirmed this association. Furthermore, consistent with the literature, we found that symptomatic patients exhibited a higher degree of stenosis (26), greater plaque burden and thickness (27,28), and a higher prevalence of vulnerable plaque features such as ulceration (29) and rim sign (18,30). These characteristics have been incorporated into various risk scores or used as independent predictors. However, despite their established biological relevance and clinical accessibility, the predictive power of these conventional metrics often plateaus. This is likely because they have a limited capacity to characterize the intrinsic, more complex pathobiological features of plaque vulnerability (e.g., inflammatory activity, microscopic necrotic cores, and IPH) and may be subject to interobserver variability. Our findings reinforce that using these conventional parameters alone is insufficient for achieving highly accurate individualized stroke risk prediction.

Compared to conventional modeling, deep learning simulates the neural network mechanism of the human brain and can automatically learn and extract features from large-scale data. It is widely used in complex pattern recognition and prediction tasks. In the field of medical image processing, deep learning has significantly reduced the need for manual intervention, improving diagnostic efficiency and accuracy. It also demonstrates excellent generalization ability and effectively handles images from different sources and of varying quality, thereby compensating for the limitations of manual assessment (31,32). Early investigations predominantly examined CNNs (12), but in recent years, with advancements in artificial intelligence technology, the Transformer model has been introduced into the field of computer vision. The Swin Transformer and DenseNet121 models in this study both achieved good classification performance, and no statistically significant difference was observed between the two deep learning models. This finding is consistent with the study by Gao et al. (33), in which a fusion model (Trans-CNN) incorporating high-resolution magnetic resonance vessel wall imaging with an attention mechanism and deep learning achieved an AUC of 0.951 in predicting the risk of stroke recurrence in patients with symptomatic intracranial atherosclerotic stenosis. Another study (34) developed a fused model incorporating Vision Transformer and radiomics to predict stroke recurrence, which also demonstrated strong performance (AUC =0.963). Together, these studies highlight the potential of deep learning to generate synthetic data that can enhance model performance, as well as its ability to capture complex image patterns that may be overlooked by traditional methods.

It is worth noting that the fusion model in this study did not achieve the best performance; rather, it was the standalone Swin Transformer model that produced the best results. However, the DeLong test indicated that its classification performance in the test set was not significantly better than that of the fusion model. We speculate that this is attributable to the relative weights of different features, as the combined model assigned considerably greater importance to the more discriminative features, specifically, those derived from deep learning. This contrast suggests that traditional features may have been subsumed or overshadowed by deep learning features, if not entirely replaced by them.

Despite the encouraging results, this study involved several limitations that should be addressed in future work. First, we employed a single-center, retrospective design with a relatively limited sample size (n=264). This might have introduced selection bias and limited the model’s generalizability to broader populations (e.g., different ethnicities and scanning protocols from different institutions). Future multicenter, prospective studies with larger cohorts are necessary to further validate the model’s universality and stability. Second, regarding image data processing, our analysis was based solely on a 2D slice at the site of maximum plaque area. Although this simplified the pipeline and focused on the most significant lesion, it inevitably neglected information regarding the plaque’s overall 3D morphology, spatial distribution, and longitudinal extent. This approach may also be sensitive to variability in slice selection, and we did not assess the robustness of the model using adjacent slices. Developing fully automated 3D plaque segmentation and analysis pipelines, coupled with 3D deep learning models for feature extraction, could provide a more comprehensive characterization of plaque properties and potentially lead to further improvements in predictive performance. Third, although we provided interpretability analyses via SHAP and Grad-CAM, offering initial insights into key features and their influence, the deep learning model remains essentially a “black box” for clinicians. Achieving complete transparency in its internal decision-making logic remains a challenge. More in-depth research on systematic interpretability is needed to establish direct correlations between the deep learning features and specific pathobiological changes, as this could foster clinical trust. Finally, we enrolled patients who underwent concurrent CTA and MRI (time interval ≤7 days), with the symptomatic group defined by positivity for acute ischemic stroke on DWI. This procedural choice placed focus on the association between plaque features and acute stroke events that had already occurred, and the model might have primarily captured imaging characteristics of recently ruptured plaques rather than prerupture “at-risk plaques”. Consequently, the generalizability of the model to primary prevention scenarios (assessing future stroke risk in asymptomatic patients) remains to be verified. Prospective studies with long-term follow-up of patients with asymptomatic plaque are needed to validate the model’s ability to identify prerupture high-risk plaques and predict incident stroke events. The model and its associated weights will be made publicly available online (https://github.com/qiyue-all/Carotid-Prediction-Model upon publication).

Conclusions

We successfully developed a multimodal model that fuses Swin Transformer-derived imaging features with clinical imaging data. This model demonstrated exceptional efficacy in predicting acute ischemic stroke risk, significantly outperforming the conventional model. These findings confirm the immense potential of deep learning, particularly the Swin Transformer architecture, in comprehensively mining the risk information embedded within carotid CTA images. Despite its limitations, this study provides a valuable direction and a solid preliminary foundation for building precise and interpretable stroke risk prediction tools.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2688/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2688/dss

Funding: This study was supported by the National Natural Science Foundation of China (NSFC) (81671663) and the Natural Science Foundation of Chongqing (NSFCQ) (cstc2014jcyjA10047).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2688/coif). The authors have no conflicts of interest to declare.

Ethical Statement: 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 and its subsequent amendments. The study was approved by the Ethics Committee of The First Affiliated Hospital of Chongqing Medical University (No. K2023-388). Informed consent for this retrospective study was waived.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. GBD 2019 Stroke Collaborators. Global, regional, and national burden of stroke and its risk factors, 1990-2019: a systematic analysis for the Global Burden of Disease Study 2019. Lancet Neurol 2021;20:795-820. [Crossref] [PubMed]
2. Baradaran H, Gupta A. Carotid Vessel Wall Imaging on CTA. AJNR Am J Neuroradiol 2020;41:380-6. [Crossref] [PubMed]
3. Cyréus P, Wadén K, Hellberg S, Bergman O, Lengquist M, Karlöf E, Buckler A, Matic L, Roy J, Marlevi D, Chemaly M, Hedin U. Atherosclerotic plaque instability in symptomatic non-significant carotid stenoses. JVS Vasc Sci 2025;6:100280. [Crossref] [PubMed]
4. Pisu F, Williamson BJ, Nardi V, Paraskevas KI, Puig J, Vagal A, de Rubeis G, Porcu M, Cau R, Benson JC, Balestrieri A, Lanzino G, Suri JS, Mahammedi A, Saba L. Machine Learning Detects Symptomatic Plaques in Patients With Carotid Atherosclerosis on CT Angiography. Circ Cardiovasc Imaging 2024;17:e016274. [Crossref] [PubMed]
5. Zhang H, Long J, Liu X, Xu W, Sun X, Zhang H, Xia X, Song C, Wang Y, Zhou D, Xu X, Zhu L, Hu C, Xu K, Meng Y. Development of a nomogram model using the dual-energy computed tomography angiography parameters of carotid plaque, the vascular lumen, and perivascular fat to predict acute stroke events. Quant Imaging Med Surg 2025;15:4947-59. [Crossref] [PubMed]
6. Zhou T, Li Y, Sun N, Hua B, Zhang M, Ma Z, Li M, Jia S. Superiority of black-blood computed tomography over conventional computed tomography angiography in the identification of soft plaques: a prospective study. Quant Imaging Med Surg 2025;15:6287-300. [Crossref] [PubMed]
7. Shi J, Sun Y, Hou J, Li X, Fan J, Zhang L, Zhang R, You H, Wang Z, Zhang A, Zhang J, Jin Q, Zhao L, Yang B. Radiomics Signatures of Carotid Plaque on Computed Tomography Angiography: An Approach to Identify Symptomatic Plaques. Clin Neuroradiol 2023;33:931-41. [Crossref] [PubMed]
8. Nies KPH, Aizaz M, van Dam-Nolen DHK, Goring TCD, Schreuder TAHCML, van Orshoven NP, Postma AA, Bos D, Hendrikse J, Nederkoorn P, van der Geest R, van Oostenbrugge RJ, Mess WH, Kooi ME. Signal intensity and volume of carotid intraplaque hemorrhage on magnetic resonance imaging and the risk of ipsilateral cerebrovascular events: The Plaque At RISK (PARISK) study. J Cardiovasc Magn Reson 2024;26:101049. [Crossref] [PubMed]
9. Nardi V, Benson C, Bois MC, Saba L, Larson AS, Özcan I, Ahmad A, Morse DW, Meyer FB, Brinjikji W, Lanzino G, Lerman LO, Savastano LE, Lerman A. Carotid Plaques From Symptomatic Patients With Mild Stenosis Is Associated With Intraplaque Hemorrhage. Hypertension 2022;79:271-82. [Crossref] [PubMed]
10. Hiremath A, Shiradkar R, Fu P, Mahran A, Rastinehad AR, Tewari A, Tirumani SH, Purysko A, Ponsky L, Madabhushi A. An integrated nomogram combining deep learning, Prostate Imaging-Reporting and Data System (PI-RADS) scoring, and clinical variables for identification of clinically significant prostate cancer on biparametric MRI: a retrospective multicentre study. Lancet Digit Health 2021;3:e445-54. [Crossref] [PubMed]
11. Wang Y, Lu W, Xu L, Xu H, Kong D. Deep learning-based ultrasound diagnostic model for follicular thyroid carcinoma. Eur Radiol 2026;36:357-66. [Crossref] [PubMed]
12. Yu Q, Ning Y, Wang A, Li S, Gu J, Li Q, Chen X, Lv F, Zhang X, Yue Q, Peng J. Deep learning-assisted diagnosis of benign and malignant parotid tumors based on contrast-enhanced CT: a multicenter study. Eur Radiol 2023;33:6054-65. [Crossref] [PubMed]
13. Saba L, Sanagala SS, Gupta SK, Koppula VK, Johri AM, Sharma AM, Kolluri R, Bhatt DL, Nicolaides A, Suri JS. Ultrasound-based internal carotid artery plaque characterization using deep learning paradigm on a supercomputer: a cardiovascular disease/stroke risk assessment system. Int J Cardiovasc Imaging 2021;37:1511-28. [Crossref] [PubMed]
14. Liu Z, Lin Y, Cao Y, Hu H, Wei Y, Zhang Z, Lin S, Guo B. Swin Transformer: Hierarchical Vision Transformer using Shifted Windows. Available online: https://openaccess.thecvf.com/content/ICCV2021/papers/Liu_Swin_Transformer_Hierarchical_Vision_Transformer_Using_Shifted_Windows_ICCV_2021_paper.pdf
15. Saba L, Cau R, Murgia A, Nicolaides AN, Wintermark M, Castillo M, et al. Carotid Plaque-RADS: A Novel Stroke Risk Classification System. JACC Cardiovasc Imaging 2024;17:62-75. [Crossref] [PubMed]
16. North American Symptomatic Carotid Endarterectomy Trial. Methods, patient characteristics, and progress. Stroke 1991;22:711-20. [Crossref] [PubMed]
17. Gupta A, Baradaran H, Kamel H, Pandya A, Mangla A, Dunning A, Marshall RS, Sanelli PC. Evaluation of computed tomography angiography plaque thickness measurements in high-grade carotid artery stenosis. Stroke 2014;45:740-5. [Crossref] [PubMed]
18. Eisenmenger LB, Aldred BW, Kim SE, Stoddard GJ, de Havenon A, Treiman GS, Parker DL, McNally JS. Prediction of Carotid Intraplaque Hemorrhage Using Adventitial Calcification and Plaque Thickness on CTA. AJNR Am J Neuroradiol 2016;37:1496-503. [Crossref] [PubMed]
19. Zhang R, Zhang Q, Ji A, Lv P, Zhang J, Fu C, Lin J. Identification of high-risk carotid plaque with MRI-based radiomics and machine learning. Eur Radiol 2021;31:3116-26. [Crossref] [PubMed]
20. Dong Z, Zhou C, Li H, Shi J, Liu J, Liu Q, Su X, Zhang F, Cheng X, Lu G. Radiomics versus Conventional Assessment to Identify Symptomatic Participants at Carotid Computed Tomography Angiography. Cerebrovasc Dis 2022;51:647-54. [Crossref] [PubMed]
21. Song JW, Phi HQ, Koneru M, Cao Q, Rubin J, Sakai Y, Ibrahim L, Zhou SE, Woo JH, Kasner SE, Saba L, Cucchiara BL. Prevalence of High-Risk CTA-Based Carotid Plaque-RADS Subtypes in Patients With Embolic Stroke of Undetermined Source. Stroke 2025;56:737-40. [Crossref] [PubMed]
22. Ding PF, Zhang HS, Wang J, Gao YY, Mao JN, Hang CH, Li W. Insulin resistance in ischemic stroke: Mechanisms and therapeutic approaches. Front Endocrinol (Lausanne) 2022;13:1092431. [Crossref] [PubMed]
23. Owolabi MO, Thrift AG, Mahal A, Ishida M, Martins S, Johnson WD, Pandian J, Abd-Allah F, Yaria J, Phan HT, Roth G, Gall SL, Beare R, Phan TG, Mikulik R, Akinyemi RO, Norrving B, Brainin M, Feigin VLStroke Experts Collaboration Group. Primary stroke prevention worldwide: translating evidence into action. Lancet Public Health 2022;7:e74-85. [Crossref] [PubMed]
24. Feigin VL, Brainin M, Norrving B, Martins S, Sacco RL, Hacke W, Fisher M, Pandian J, Lindsay P. World Stroke Organization (WSO): Global Stroke Fact Sheet 2022. Int J Stroke 2022;17:18-29. [Crossref] [PubMed]
25. Zhu J, Jiang Y. Editorial: the role of diabetes in the pathophysiology and prognosis of ischemic stroke. Front Endocrinol (Lausanne) 2023;14:1207537. [Crossref] [PubMed]
26. Han L, Li Z, Wang J, Gao Y, Zhang L, Han S, Li L, Yang Z, Ma X, Liu H, Peng S, Zhao H, Cui H. Association Between Intracranial Plaque Characteristics and Stroke Recurrence Risk Stratified by Stenosis Severity. Acad Radiol 2025;32:6205-15. [Crossref] [PubMed]
27. Brunner G, Virani SS, Sun W, Liu L, Dodge RC, Nambi V, Coresh J, Mosley TH, Sharrett AR, Boerwinkle E, Ballantyne CM, Wasserman BA. Associations Between Carotid Artery Plaque Burden, Plaque Characteristics, and Cardiovascular Events: The ARIC Carotid Magnetic Resonance Imaging Study. JAMA Cardiol 2021;6:79-86. [PubMed]
28. Yoon Y, Lee DH, Kang DW, Kwon SU, Kim JS. Single subcortical infarction and atherosclerotic plaques in the middle cerebral artery: high-resolution magnetic resonance imaging findings. Stroke 2013;44:2462-7. [Crossref] [PubMed]
29. Baradaran H, Al-Dasuqi K, Knight-Greenfield A, Giambrone A, Delgado D, Ebani EJ, Kamel H, Gupta A. Association between Carotid Plaque Features on CTA and Cerebrovascular Ischemia: A Systematic Review and Meta-Analysis. AJNR Am J Neuroradiol 2017;38:2321-6. [Crossref] [PubMed]
30. Altaf N, MacSweeney ST, Gladman J, Auer DP. Carotid intraplaque hemorrhage predicts recurrent symptoms in patients with high-grade carotid stenosis. Stroke 2007;38:1633-5. [Crossref] [PubMed]
31. Lee SB, Cho YJ, Hong Y, Jeong D, Lee J, Kim SH, Lee S, Choi YH. Deep Learning-Based Image Conversion Improves the Reproducibility of Computed Tomography Radiomics Features: A Phantom Study. Invest Radiol 2022;57:308-17. [Crossref] [PubMed]
32. Yaqub M, Jinchao F, Arshid K, Ahmed S, Zhang W, Nawaz MZ, Mahmood T. Deep Learning-Based Image Reconstruction for Different Medical Imaging Modalities. Comput Math Methods Med 2022;2022:8750648. [Crossref] [PubMed]
33. Gao Y, Li ZA, Xie BC, Wang WP, Sun YC, Wei ZQ, Zhai XY, Zhao QY, Han L, Du X, Wang J, Zhang P, Yan RF, Li YD, Cui HK. Deep learning network based on high-resolution magnetic resonance vessel wall imaging combined with attention mechanism for predicting stroke recurrence in patients with symptomatic intracranial atherosclerosis. Quant Imaging Med Surg 2025;15:2929-43. [Crossref] [PubMed]
34. Gao Y, Li Z, Zhai X, Zhang G, Zhang L, Huang T, Han L, Wang J, Yan R, Li Y, Zhao H, Zhao Q, Wei Z, Xie B, Sun Y, Zhao J, Cui H. MRI-based habitat radiomics combined with vision transformer for identifying vulnerable intracranial atherosclerotic plaques and predicting stroke events: a multicenter, retrospective study. EClinicalMedicine 2025;82:103186. [Crossref] [PubMed]