AI-driven radiomics in vascular disease: a narrative review of image analysis and clinical translation

Introduction

Imaging techniques are critical tools in the diagnosis and management of vascular diseases, enabling the detailed visualization of vascular morphology and the early detection of pathological changes, even in asymptomatic individuals (1). Coronary computed tomography angiography (CCTA) is routinely used to exclude coronary artery stenosis, and characterize vessel wall architecture and plaque morphology with high spatial resolution (2). Intravascular ultrasound (IVUS), the current gold standard for vascular lesion evaluation, provides cross-sectional images to assess calcification and vessel wall integrity (3). Due to its non-invasive nature and three-dimensional (3D) imaging capability, time-of-flight magnetic resonance angiography (TOF-MRA) offers diagnostic performance comparable to that of conventional angiography and is widely applied in clinical practice (4). These modalities provide complementary insights into vascular structure and pathology, forming the basis for personalized risk stratification and decision-making.

Imaging techniques remain constrained by factors such as artifacts, limited resolution, and low signal-to-noise ratio (SNR), often resulting in suboptimal image quality. Such variability in image acquisition and interpretation challenges the accuracy and reproducibility of clinical assessments (5). These discrepancies can introduce noise into ground-truth labels, compromising artificial intelligence (AI) model training and validation. Interpreting vascular images requires specialized expertise and places high demands on clinicians. Inconsistencies across imaging platforms further limit reproducibility and standardization in workflows.

The ability of current imaging techniques to capture functional information, particularly radiomic, anatomical, and histological features, remains limited. AI enables more efficient and scalable analysis, enhancing expert interpretation (6).

Radiomics, a core AI-driven approach in medical imaging, enables the high-throughput extraction of quantitative image features for analysis (7). The integration of AI technology offers a potential solution to the limitations of conventional image analysis methods (Figure 1). This review summarizes current radiomics applications in vascular disease diagnosis, prediction, and treatment, and discusses the associated challenges and future impacts. We present this article in accordance with the Narrative Review reporting checklist (available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-531/rc).

Figure 1 Clinical management of vascular disease using AI applications. This figure delineates the comprehensive integration of AI across various stages of the clinical pathway for vascular diseases. The process commences with diagnosis, where AI algorithms assist in the precise identification of plaques and the accurate detection of stenosis. Subsequently, in the treatment phase, AI applications extend to radiogenomics, supporting personalized therapeutic approaches and guiding endovascular procedures. Finally, AI significantly contributes to prognosis, aiding in survival prediction and facilitating the monitoring of disease progression over time. Created in BioRender. Fu Y [2026]. https://BioRender.com/4ghcp7f. AI, artificial intelligence.

Methods

A comprehensive literature search was conducted of the PubMed database to retrieve relevant English-language articles published from 2015 to 2025. The search keywords included “artificial intelligence”, “machine learning”, “deep learning”, “radiomics”, “vascular imaging”, “plaque”, and “calcification”. Large-scale multicenter studies and pivotal clinical trials were prioritized. The search strategy is summarized in Table 1.

Results

AI development

AI originated in the mid-20th century, with the term introduced by John McCarthy in 1956 (8). However, limited computational power led to a period of stagnation during the 1970s, known as the “AI winter” (9). The 1990s saw a revival in AI driven by machine learning (ML), exemplified by milestones such as IBM’s Deep Blue defeating Garry Kasparov in chess in 1997 (10). In recent decades, advances in deep learning (DL) and the availability of large-scale data have positioned AI as a transformative force in various sectors, including healthcare, finance, and autonomous driving (11).

ML and DL

Recent advancements in computer vision have facilitated the integration of ML and DL into medical imaging, allowing algorithms to extract a large number of features and correlate them with clinical outcomes (12) (Table 2). Commonly used ML algorithms include supervised methods [e.g., decision trees, Bayesian classifiers, and support vector machines (SVMs)], unsupervised methods (e.g., k-means clustering), and semi-supervised approaches (13) (Figure 2).

Figure 2 Classification of ML algorithms. This hierarchical diagram presents a comprehensive classification of common ML algorithms, broadly categorized into two principal paradigms: traditional ML and DL. Traditional ML methods include a range of algorithms, including decision trees, SVMs, and various ensemble learning techniques such as random forest and boosting. Conversely, DL represents a more advanced subset of AI, characterized by multiple interconnected layers, and includes complex architectures like CNNs and RNNs. AI, artificial intelligence; CNN, convolutional neural network; DL, deep learning; GNN, graph neural network; KNN, K-nearest neighbor; LSTM, long short-term memory; ML, machine learning; RNN, recurrent neural network; SVM, support vector machine; VGG, Visual Geometry Group.

As a subset of ML, DL is particularly effective at handling high-dimensional data and identifying complex patterns. Among DL models, convolutional neural networks (CNNs) are widely applied for image segmentation and reconstruction. A standard CNN architecture includes convolutional, pooling, and fully connected layers, optimized to process multi-dimensional array data (14) (Figure 3). In medical imaging, generative adversarial networks (GANs) enhance reconstruction quality and generate synthetic data to improve generalization and address privacy concerns (15), while encoder-decoder architectures facilitate automated vascular segmentation to increase diagnostic accuracy and efficiency (16).

Figure 3 Basic architecture of a CNN. The network begins with an input layer, which is subsequently processed through one or more sequences of Conv layers and max pooling layers. Conv layers are responsible for extracting features from the input data, while max pooling layers reduce the dimensionality and provide translation invariance. Following these feature extraction stages, the data are typically transformed into a one-dimensional vector by a flatten layer, which then feeds into one or more fully FC layers. These FC layers perform high-level reasoning and pattern recognition, ultimately leading to a final output layer that typically provides classification probabilities or regression values. Created in BioRender. Fu Y [2026]. https://BioRender.com/nbjgda1. CNN, convolutional neural network; Conv, convolutional; FC, fully connected; max, maximum; ReLU, rectified linear unit.

ML and DL have significantly advanced radiologists’ image interpretation and workflow efficiency. AI can significantly reduce the reading time and enhance the diagnostic throughput of radiologists, who primarily interpret imaging studies and generate diagnostic reports (17). Trained models have been shown to reduce inter- and intra-observer variability, increasing confidence in imaging assessments and improving consistency across multicenter studies (18). In addition, AI algorithms can automatically extract voxel-level information and identify relevant imaging features, facilitating the early diagnosis of vascular diseases and potentially improving patient outcomes (19) (Figure 4).

Figure 4 General workflow of an AI-based medical image analysis pipeline. The process begins with image acquisition and preprocessing, where raw medical images are obtained from various modalities and prepared for subsequent analysis through steps like noise reduction, bias correction, and standardization. This is followed by segmentation, which involves delineating specific anatomical structures, regions of interest, or abnormalities in the images (e.g., vessels). Next, feature extraction and selection is performed, where relevant quantitative features (e.g., textural, morphological and intensity-based) are derived from the segmented regions, and optimal features are selected to build predictive models. Finally, the selected features are used for modeling and validation, where machine learning or deep learning models are trained on the data and rigorously evaluated for their performance in tasks such as diagnosis, prognosis, or treatment response prediction. Created in BioRender. Fu Y [2026]. https://BioRender.com/wn0h9ki. AI, artificial intelligence; AUC, area under the curve; ROC, receiver operating characteristic; S.D., standard deviation; SVM, support vector machine.

To capture current research trends, a review of studies from 2015 to present on ML/DL in vascular diseases revealed that CNNs, especially the U-Net architecture, are the most widely used models for image segmentation. Among traditional ML algorithms, SVMs and random forests remain common in radiomics-based predictive modeling. These trends indicate a preference for models that offer a balance of computational efficiency and interpretability, while emerging architectures like Transformers and graph neural networks (GNNs) show promise for handling more advanced tasks (Figure 5).

Figure 5 Distribution of AI methodologies and specific algorithms/architectures in reviewed studies. This figure presents the statistical distribution of different AI methodologies and their specific components as observed in the analyzed research studies. (A) Overall distribution of methods showing the primary AI approaches used in the analyzed studies; notably, DL constituted the largest proportion (63.93%), followed by traditional ML (29.51%), and hybrid methods (6.56%). (B) Detailed distribution of traditional ML algorithms showing the specific algorithms employed in the traditional ML paradigm. (C) Detailed distribution of DL architectures highlighting the prevalence of various architectures, among which, U-Net and its variants were overwhelmingly dominant (66.67%). AI, artificial intelligence; ANN, artificial neural network; CNN, convolutional neural network; DL, deep learning; ML, machine learning; SVM, support vector machine.

Imaging enhancement

In the diagnosis and management of vascular diseases, AI, particularly ML and DL, has emerged as a pivotal technology for enhancing medical image quality. It has been applied across ultrasound, computed tomography (CT), and magnetic resonance imaging (MRI) to enhance resolution, increase the SNR, reduce the acquisition time, and lower computational demands, thereby addressing the inherent limitations of conventional imaging modalities (20,21).

Recent research has explored various strategies to produce higher-quality images that support better interpretation and clinical decision-making (4,20-36) (Table 3). AI-driven image enhancement not only improves visual quality but also accelerates image acquisition and enables more robust quantitative analysis, supporting early diagnosis and personalized treatment planning for vascular diseases (15). The following sections outline the clinical applications of these technologies across various imaging modalities, focusing on four key areas: denoising, resolution enhancement, motion compensation, and direct reconstruction.

Denoising

AI-based denoising techniques outperform traditional methods (e.g., median filtering and wavelet transforms) in vascular imaging by effectively removing noise while preserving critical vascular features. This capability enhances the visualization of vascular boundaries and improves diagnostic accuracy for conditions like stenosis and thrombosis, overcoming the limitations of noise-induced degradation that hamper traditional approaches (37).

Hemodynamic analysis provides essential information about vascular function by evaluating parameters such as blood flow velocity, pressure gradients, and shear stress, which are fundamental for assessing vascular health and the progression of disease. Four-dimensional flow MRI (4D flow MRI) enables the visualization and measurement of these parameters but is often limited by noise-induced degradation. To address this issue, a CNN was trained with computational fluid dynamics (CFD)-simulated velocity fields to enhance low-resolution 4D flow MRI data. Subsampled MRI scans served as input, while CFD-derived velocity maps acted as reference labels, enabling the CNN to effectively suppress noise and restore fine vascular flow details. Compared to the 24-minute high-resolution reference scan, the root mean square error (RMSE) decreased from 2.55 to 1.49 for 12-minute scans and from 4.06 to 1.64 for 6-minute scans, enabling more accurate and reliable hemodynamic assessment (25).

In addition to supervised DL approaches, researchers have investigated unsupervised learning strategies to enhance medical image quality. Compressed sensing MRI has been employed to reduce the lengthy acquisition time of high-resolution proton density-weighted MRI for intracranial vessel wall imaging, but often at the cost of image quality. In a study of 14 healthy volunteers, a U-Net-based model integrated with CycleGAN for unsupervised training yielded encouraging results. The output images demonstrated reduced noise, a significantly increased SNR (P<0.05), and enhanced reproducibility of radiological features. Further, the model improved the consistency of both texture and wavelet-based radiomic features (21). However, the study also noted that the noise reduction achieved by unsupervised learning was inferior to that of supervised approaches. This underscores a common limitation of unsupervised learning. The trade-off between reproducibility and output quality must be addressed to advance the clinical utility of unsupervised learning methods.

Resolution enhancement

Image resolution is crucial for the detection and diagnosis of vascular lesions. High-resolution images enhance the visualization of tissue structures, boundaries, and pathological features, thereby improving diagnostic accuracy and reliability (38). However, achieving such quality typically requires advanced hardware and considerable computational resources, which limits the broader clinical adoption of such imaging techniques (33). DL has emerged as a promising solution to these limitations.

Zhou et al. developed a novel two-stage GAN model that incorporates a U-Net-based pre-generator module. The architecture aims to reconstruct structural details, suppress speckle noise, and enhance overall image quality. The model improved low-quality carotid ultrasound images, increasing the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) by 10.65 dB and 0.39, respectively (26). These enhancements significantly improved both image quality and the preservation of structural information, thereby supporting more accurate diagnostic decisions.

In another carotid artery study, researchers integrated radial sampling with a U-Net-based DL model, enabling ungated quiescent interval slice-selective (QISS) MRA in under three minutes while substantially improving image quality (27). This method achieves a practical balance between acquisition speed and image quality, underscoring its potential for real-world clinical deployment.

Motion correction

Motion correction is critical in medical imaging, particularly in cardiovascular applications, where cardiac and respiratory motion can introduce artifacts that degrade image quality and hinder diagnostic accuracy (39). ML, particularly DL, has emerged as a powerful solution to address these challenges by significantly improving image quality and processing efficiency. For instance, the Respiratory Motion Estimation network (RespME-net) framework uses unsupervised DL to estimate non-rigid respiratory motion in whole-heart coronary magnetic resonance (MR) images (29). By rapidly predicting 3D non-rigid motion fields in coronary MRA, it achieves image quality comparable to that of traditional registration methods, while reducing the motion estimation time from several minutes to seconds—a 20-fold improvement—highlighting the precision and efficiency of DL.

In ultrasound imaging, Park et al. introduced the Deep Learning Boundary Detection and Compensation (DL-BDC) technique for vascular boundary segmentation and wall motion compensation (32). This approach substantially improved both ultrasound image quality and blood flow quantification accuracy. In intravascular imaging, deep features have been applied for intra-slice motion correction in optical coherence tomography (OCT) images. By leveraging CNNs to extract and align deep features, the method has been shown to be effective in targeted applications, although its scope remains relatively limited (40). DL also facilitates the automation of cardiac rest period detection for calculating trigger delays in coronary MRA. This automated workflow enhances reproducibility and streamlines the process compared to manual decision-making (30).

In summary, ML has driven substantial advances in motion correction for vascular imaging. These developments underscore the transformative potential of DL in clinical imaging to improve image quality, shorten the processing time, and enhance diagnostic accuracy.

Direct reconstruction

DL has revolutionized direct vascular reconstruction by addressing the inherent trade-offs of traditional imaging methods (e.g., compressed sensing), which often struggle to balance image quality and acquisition efficiency (41). Through the integration of unsupervised and advanced supervised learning strategies, recent studies have markedly enhanced both the accuracy and efficiency of vascular imaging.

Notably, the application of unsupervised DL methods has eliminated the need for paired datasets. Chung et al. proposed a CycleGAN-based framework incorporating optimal transport theory to expedite 3D TOF-MRA acquisition. Their model outperformed conventional compressed sensing techniques and achieved results comparable to those of supervised methods, highlighting its potential in data-scarce scenarios (4). In the context of TOF-MRA, supervised DL approaches have further expanded the capabilities of vascular imaging. Jun et al. developed Deep Parallel Imaging network (DPI-net), a deep multi-stream CNN, to reconstruct 3D multi-channel MRA from undersampled data. DPI-net showed superior performance in preserving vascular signals in both axial slices and maximum intensity projection images, outperforming traditional parallel imaging and other DL models (35).

Another notable innovation is the integration of super-resolution techniques with DL to enhance vascular imaging. Küstner et al. presented a DL-based super-resolution framework for 3D isotropic coronary MRA, achieving 16-fold spatial resolution enhancement while enabling free-breathing acquisition in under a minute. By incorporating GANs and motion compensation, the model enhanced vascular sharpness and length without compromising diagnostic accuracy (31). Collectively, these approaches highlight the versatility of DL in addressing key challenges in vascular reconstruction, paving the way for efficient and high-quality imaging across various modalities.

Pathological analysis

Vascular lesions affect the cardiovascular, cerebrovascular, and peripheral vascular systems, contributing to conditions such as myocardial infarction, stroke, and peripheral artery disease (42). Early detection and accurate characterization, particularly in identifying atherosclerotic plaques, calcification, vascular stiffness, and stenosis, are essential for effective risk assessment and intervention (43). Understanding vascular pathology requires both the identification of structural abnormalities and the evaluation of their functional consequences and progression. While conventional imaging provides essential morphological data, it often fails to capture the subtle textural heterogeneity and functional dynamics of lesions (44).

Radiomics addresses these limitations by extracting high-dimensional quantitative features (texture, shape, and intensity) from images, enabling superior differentiation between stable and high-risk plaques, refined calcification classification, and comprehensive stiffness assessment (45). Further, the integration of radiomics with CFD and 4D flow MRI enables non-invasive hemodynamic evaluation, providing insights into the functional impact of vascular stenosis and supporting risk stratification (46). These advancements collectively enhance diagnostic precision, support the predictive modeling of disease progression, and pave the way for personalized treatment strategies. The subsequent sections detail AI applications in vascular lesion analysis across imaging modalities.

Plaque

The development of atherosclerotic plaques, characterized by lipid accumulation, chronic inflammation, and extracellular matrix remodeling, constitutes a central pathological process in cardiovascular diseases and remains a major focus of vascular lesion research (47). Although CCTA has long been considered the gold standard for evaluating coronary artery anatomy, its limited ability to assess vessel wall and plaque morphology has led to the adoption of intravascular imaging techniques such as digital subtraction angiography (DSA) (47). Recent advances in AI have expanded the capabilities of non-invasive imaging modalities, enabling the more accurate evaluation of plaque burden and morphological features.

Recent studies have primarily focused on leveraging AI for plaque detection and segmentation. In ultrasound imaging, an automated framework incorporating a dual-attention U-Net was developed for simultaneous vessel and plaque segmentation and tracking, and a multi-stream similarity learning network was developed for precise target tracking and segmentation between frames. The framework achieved a Dice similarity coefficient (DSC) of 0.83 (48). In IVUS, Blanco et al. employed a multi-frame CNN and Gaussian process regressor to segment the vessel wall, lumen, and plaque. The Jaccard index for both the vessel lumen and wall exceeded 0.9 (3).

Radiomics-based analysis has further enabled the detailed characterization of high-risk plaques. Lin et al. demonstrated that radiomics-based precision phenotyping using CCTA effectively distinguishes between the culprit and highest-grade non-culprit plaques in patients with myocardial infarction. The two plaque types exhibited distinct radiomic signatures, with texture and geometric features serving as key discriminators. Notably, culprit plaques were characterized by larger volumes of non-calcified and low-density non-calcified components, as well as a higher prevalence of adverse morphological features compared to non-culprit plaques. These findings highlight the potential of radiomics in the non-invasive identification of vulnerable plaques and the stratification of high-risk patients (49). Zhang et al. investigated the association between carotid atherosclerotic plaques and ischemic stroke, successfully distinguishing between symptomatic and asymptomatic plaques. The proposed radiomics model achieved an area under the curve (AUC) of 0.988, markedly surpassing the performance of conventional approaches (50).

However, the field has recently moved beyond small-scale segmentation studies to large-scale multicenter clinical validations using AI-driven quantitative CT (AI-QCT). The CLARIFY study showed that AI-QCT provided superior reproducibility and precision in quantifying stenosis severity and plaque burden compared to conventional visual assessment by radiologists. The AI-based evaluation significantly reduced inter-reader variability, offering a standardized method for grading stenosis (51). Further, AI has demonstrated robust performance in analyzing complex lesions in high-risk populations. For instance, in patients with end-stage renal disease, AI-augmented CCTA successfully quantified plaque volumes and distinct subtypes, providing critical risk assessment in a population where traditional imaging is often challenged by heavy calcification (52).

Crucially, AI facilitates the longitudinal monitoring of disease progression. A recent serial analysis using AI-CCTA tracked coronary artery disease over a 13-year period, successfully identifying significant changes in total, calcified, and non-calcified plaque volumes. The ability to quantify minute temporal changes in plaque composition represents a significant advancement for monitoring therapeutic efficacy and disease trajectory (53).

These developments reflect a paradigm shift in plaque research, from basic detection toward advanced compositional analysis and risk classification. Such progress offers robust tools for improving the precision of cardiovascular diagnosis and management, potentially enhancing clinical outcomes.

Calcification

Vascular calcification, characterized by the in situ deposition of calcium and phosphate in the vascular wall, commonly occurs in patients with chronic kidney disease (CKD), diabetes, and atherosclerosis. As a hallmark of arterial dysfunction, it is strongly associated with increased risks of cardiovascular and cerebrovascular events. The distribution and severity of calcification are essential for evaluating disease progression and predicting clinical outcomes.

Coronary artery calcification (CAC) scoring systems, including the Agatston score, volume score, and mass score, are commonly used in clinical practice to quantify calcification and stratify risk. CAC scoring has been shown to independently predict major adverse cardiovascular events, surpassing the prognostic accuracy of traditional risk factors (54). High CAC scores reflect a greater calcific burden and are associated with higher risks of myocardial infarction, stroke, and all-cause mortality. However, traditional manual measurement methods are subject to variability, poor reproducibility, and dependence on imaging quality, limiting their applicability in large-scale screening and precision medicine (55). The integration of AI has revolutionized calcification assessment by enabling automated identification, precise quantification, and individualized risk prediction, while also improving efficiency and standardization.

AI has made substantial progress in calcification detection through DL models, which enable automated segmentation and identification. Zhang et al. developed a DL framework incorporating multi-view shape constraints for precise CAC detection, facilitating the comprehensive quantification of both total and vessel-specific calcification burden (56). This approach minimizes human error and improves consistency across diverse imaging datasets. Similarly, Sartoretti et al. applied a 3D U-Net-based DL tool to positron emission tomography (PET)/CT scans, enabling CAC detection even in non-gated CT images and offering a viable alternative for patients ineligible for conventional cardiac CT (57).

More importantly, recent pivotal trials have demonstrated that AI can outperform traditional diagnostic metrics in detailed risk stratification and ischemia prediction. The landmark CONFIRM2 registry highlighted the prognostic power of AI-QCT in a large multicenter cohort. The study found that AI-quantified non-calcified and calcified plaque volumes were superior to traditional clinical risk scores in predicting all-cause mortality and myocardial infarction. Notably, this study revealed that for equivalent amounts of plaque, women faced a higher relative risk than men, a nuance previously under-recognized by conventional scoring (58).

AI also bridges the gap between anatomical imaging and functional ischemia. The CREDENCE trial provided compelling evidence that the AI analysis of plaque characteristics and vascular morphology predicted myocardial ischemia with high accuracy compared to invasive fractional flow reserve and single-photon emission CT. By integrating features such as lumen volume, plaque burden, and vessel geometry, AI models can identify functionally significant stenosis that may require intervention, reducing the need for invasive diagnostic procedures (59). These studies underscore the capability of AI to enhance calcification assessment across diverse imaging modalities, improving the comprehensiveness and accessibility of such assessments.

In calcification quantification, traditional Agatston scoring is affected by scanning parameters, motion artifacts, and subjective manual annotation. AI-based automated analysis enhances the accuracy and reproducibility of calcification quantification. Yuan et al. developed a DL model based on a Residual (2+1) D CNN [R (2+1) D CNN] to predict CAC scores from transthoracic echocardiography videos (60). This non-invasive approach demonstrated high discriminative performance in external validation, suggesting that AI can extract calcification-related features from non-CT imaging modalities, providing an alternative for patients ineligible for CT scans. Additionally, Dobrolinska et al. compared AI-derived CAC scores with manual and visual assessments on low-dose CT, and found that while the AI method showed promise, visual assessment achieved the highest agreement with gold standard CAC scores. Thus, AI models need to be further refined to accommodate diverse scanning protocols and imaging conditions (61).

Beyond detection and quantification, AI plays a pivotal role in risk stratification by linking calcification metrics to cardiovascular event risk prediction. Greenland et al. confirmed the predictive value of CAC scores in diverse populations, emphasizing their role in guiding primary prevention strategies, including lipid-lowering and anti-platelet therapies (54). Johri et al. further applied recurrent neural networks (RNNs) and long short-term memory (LSTM) networks to integrate carotid ultrasound data, intraplaque neovascularization, and conventional risk factors, achieving an AUC of 0.99 for multi-class coronary artery disease classification (62). These findings underscore the potential of AI in leveraging multimodal data for refined cardiovascular risk stratification.

Despite these advancements, several challenges persist. Variability in imaging protocols, data sources, and feature extraction methods hampers model generalizability. In addition, most current AI models rely on static imaging; however, calcification is an inherently dynamic and progressive process. Future research should integrate temporal modeling, advanced radiomics, whole-genome analysis, and CFD-based simulations to enhance risk prediction and advance precision cardiovascular medicine (63,64).

Vascular stiffness

Vascular stiffness is a hallmark of vascular aging, reflecting the progressive loss of arterial elasticity and functional capacity. Traditional assessment methods, such as pulse wave velocity (PWV) and ultrasound imaging, are effective but often limited by operator dependency and resource intensity (65). AI has introduced efficient and automated alternatives to address these limitations.

One study enhanced ARTSENS, a non-invasive arterial stiffness measurement system, by integrating a deep neural network pipeline. The model achieved 99.8% frame classification accuracy and reduced the elastic modulus estimation error by 9.3%, operating in real time with processing speeds below 20 ms per frame (66). Another study employed unsupervised learning to estimate vascular wall displacement, resulting in enhanced spatial resolution and improved accuracy in PWV assessment across both phantom and in vivo datasets (67). These AI-driven innovations streamline stiffness assessment and provide scalable, accurate tools for both clinical and research applications.

Vascular stenosis

As a hallmark of ischemic disease, vascular stenosis, defined by arterial narrowing that limits perfusion, remains a central focus of clinical diagnosis (68). Although traditional imaging techniques such as angiography and Doppler ultrasound are widely used to detect stenosis, they are time-consuming and highly dependent on operator skill (69). AI technologies are driving a paradigm shift toward more efficient and automated diagnostic solutions.

A multicenter, multi-vendor study implemented a CNN-based system for detecting and grading coronary stenosis, integrating imaging features with ML techniques. The use of gradient boosting decision trees yielded an AUC of 0.910 for stenosis grading, with a sensitivity and specificity of 84.1% and 95.7%, respectively (70). This automated approach significantly reduced reliance on manual interpretation, offering a streamlined and consistent alternative for coronary diagnostics. However, the uneven distribution of plaque and stenosis types in the training data limits the generalizability of the model, highlighting the need for more diverse datasets to improve model performance.

In renal artery imaging, researchers used a double head region-based CNN model to optimize sampling position selection in color Doppler sonography. Achieving an average accuracy of 88.5% across multiple image categories, this method improved sonographic efficiency and delivered reliable diagnostic performance for renal artery stenosis (71).

These AI-driven innovations reduce dependence on operator expertise and minimize inter-observer variability, enabling faster and more accurate stenosis detection. Through advanced automation and enhanced consistency, AI technologies have the potential to transform vascular diagnostics and support improved clinical outcomes.

Hemodynamics

Hemodynamic analysis provides essential insights into the functional implications of vascular lesions, complementing morphological assessments. By linking structural abnormalities to corresponding hemodynamic alterations, hemodynamic research enables a comprehensive understanding of disease progression and therapeutic strategies (72). AI is revolutionizing this field by introducing precise and efficient methodologies that integrate seamlessly into clinical workflows.

A notable advancement is the application of a 3D nnU-Net-based DL pipeline for 4D flow cardiac MRI. In a study by Garrido-Oliver et al., the pipeline achieved a DSC of 0.949 for automated aortic segmentation and demonstrated strong agreement with manual assessments for systolic flow reversal ratio and wall shear stress. By automating complex 4D flow analyses, this approach enhances the accessibility and clinical applicability of advanced imaging modalities (73).

In ultrasound imaging, He et al. developed a one-dimensional CNN for blood flow velocimetry in the carotid artery. The model outperformed traditional methods such as high-pass filtering and singular value decomposition, yielding lower normalized root mean square errors (NRMSE) and improved accuracy in bidirectional flow measurements (74). This approach highlights the ability of AI to enhance real-time vascular flow analysis.

Additionally, Bratt et al. employed a U-Net-based model for the automated quantification of aortic flow from phase velocity-encoded coronary MR, achieving a correlation coefficient of 0.99 with manual measurements (75). Similarly, Cai et al. used the eXtreme Gradient Boosting (XGBoost) algorithm to predict cerebral perfusion status from internal carotid artery flow, achieving an AUC of 87.08% (76). These findings show the versatility of AI in advancing hemodynamic analysis across diverse imaging modalities. AI bridges imaging data with actionable clinical insights, improving diagnostic efficiency, enabling personalized treatment strategies, and furthering our understanding of vascular and systemic hemodynamics.

Clinical application

By extracting high-dimensional quantitative features, AI-driven radiomics is transforming vascular imaging, enabling objective disease characterization, risk stratification, and treatment guidance (77). Applications include diagnosing atherosclerosis, quantifying calcification, predicting plaque vulnerability, and evaluating stiffness—all essential for risk stratification (78). ML optimizes feature selection, enhances predictive modeling, and automates workflows, reducing inter-observer variability (79). Integrating radiomics with clinical and genomic data fosters a comprehensive understanding of vascular disease and advances precision vascular medicine (80). The following sections review current applications of AI-driven radiomics in vascular imaging, emphasizing its clinical utility and future potential.

Diagnosis

Clinicians diagnose diseases by integrating information on patient history, presenting symptoms, physical examination findings, and auxiliary test results, including traditional imaging results (81). The accuracy, specificity, and sensitivity of imaging modalities vary depending on the disease context (82). AI advancements have enhanced diagnostic performance, providing imaging solutions that are more cost-effective, non-invasive, and associated with lower procedural risks. Moreover, ML mitigates observer variability in image interpretation, standardizing assessments to improve consistency and reliability (18) (Figure 6).

Figure 6 Radiomics application in vascular disease diagnosis. AI-driven radiomics offers a comprehensive framework for enhancing clinical diagnosis. Ultimately, AI-driven radiomics models provide objective, data-driven insights that significantly support clinicians to make more precise and informed diagnostic judgements. Created in BioRender. Fu Y [2026]. https://BioRender.com/yoy8vvi. AI, artificial intelligence; AUC, area under the curve; ROC, receiver operating characteristic.

Chen et al. developed a cascaded GAN framework to synthesize high-resolution CTA images from standard CT scans, thereby improving diagnostic accuracy. The generated images provided detailed vascular visualization, enabling the accurate identification of carotid dissection without extensive manual input. This method streamlined diagnostic workflows while preserving image quality (83).

Claux et al. developed a U-Net-based model to detect intracranial aneurysms in TOF-MRA scans. The model had a sensitivity of 78%, effectively detecting aneurysms of various sizes and at different locations. Its automated segmentation and detection capabilities highlight the efficiency of DL in completing complex 3D vascular imaging tasks (84).

Wu et al. introduced the deep morphology-aided diagnosis (DeepMAD) network for vessel wall segmentation and plaque characterization on black-blood vessel wall MRI. The model achieved high segmentation accuracy (DSC >0.95) and had a diagnostic AUC of 0.95. By automating vessel wall boundary identification and plaque characterization, DeepMAD offers a practical tool to support radiologists in carotid artery disease management (85).

These AI-powered approaches underscore the potential of ML to enhance diagnostic precision in cardiovascular and cerebrovascular imaging. By automating complex tasks, they reduce variability, and improve the efficiency of disease detection and assessment.

Prediction of adverse outcomes

AI has shown considerable promise in predicting adverse clinical outcomes by integrating radiomics, ML, and DL with various imaging modalities and clinical parameters (86) (Figure 7). Numerous studies have explored AI-driven approaches for predicting post-treatment outcomes, disease progression, and cardiovascular risk stratification.

Figure 7 Integration of radiomics and clinical variables for vascular disease prognosis. By integrating imaging features with clinical variables and applying machine learning algorithms (e.g., logistic regression) precise risk stratification and prognostic assessment for vascular diseases can be achieved. Created in BioRender. Fu Y [2026]. https://BioRender.com/b3ko4h0.

For stroke and endovascular treatment (EVT), Zhang et al. and Luo et al. developed radiomics models based on diffusion-weighted imaging to predict outcomes in patients with acute basilar artery occlusion undergoing EVT. Using SVM classifiers and least absolute shrinkage and selection operator (LASSO) for feature selection, their models achieved AUC values of 0.870 and 0.897 in training cohorts, respectively, demonstrating strong predictive performance for functional outcomes and futile recanalization (87,88). These studies underscore the utility of radiomics in identifying the patients most likely to benefit from EVT, thereby optimizing treatment selection.

In arterial disease risk prediction, Feng et al. and Lee et al. applied radiomics and conventional plaque parameters from CCTA to predict plaque progression and the formation of new plaques (89,90). Their models, which incorporated LASSO and Cox proportional hazards regression, demonstrated superior predictive accuracy compared with traditional risk factor-based assessments. Similarly, Johri et al. enhanced predictive performance by incorporating intraplaque neovascularization features—extracted during carotid ultrasound acquisition—into DL models (an RNN and LSTM), achieving an AUC of 0.99 for coronary artery disease classification (62). To predict the stroke risk of diabetic patients, Liu et al. developed a carotid ultrasound radiomics-based nomogram that combined clinical risk factors with radiomic features, achieving an AUC of 0.898 (91). This approach demonstrated the feasibility of non-invasive stroke risk stratification in high-risk populations.

By integrating multimodal imaging with ML, AI enhances the prediction of adverse outcomes. This AI-driven risk stratification improves patient selection, monitoring, and personalized treatment strategies. However, limited external validation, heterogeneous datasets, and a lack of standardized protocols remain key barriers to the clinical translation of AI-driven risk stratification tools (7).

Radiogenomics

In recent years, AI has achieved significant advances in the field of radiogenomics. By combining DL with genetic analysis, researchers have determined the genetic basis of cardiovascular conditions and achieved accurate risk prediction using polygenic risk scores (92) (Figure 8).

Figure 8 Integration of radiomics and genomics for clinical application in vascular disease. This figure illustrates the integration of imaging features and gene information in vascular disease. Imaging features derived from the cardiovascular system with atherosclerotic plaque were found to be correlated with gene information. This correlation analysis identified genes associated with radiomic features-associated genes, which were then used to generate PRSs from SNPs. The fusion of radiomic features and genomic information, known as radiogenomics, enables accurate clinical outcome prediction and supports the implementation of precision medicine. Created in BioRender. Fu Y [2026]. https://BioRender.com/kx6nso7. PRS, polygenic risk score; SNP, single nucleotide polymorphism.

Pirruccello et al. were among the first to apply DL to automate cardiac MRI analysis to predict aortic disease risk. They employed a U-Net-based DL model to accurately measure aortic diameters, which were then integrated with genome-wide association studies (GWAS) and transcriptome-wide association studies (TWAS) to identify genetic loci linked to ascending and descending aortic diameters. Further, polygenic risk scores revealed strong associations between the incidence of aortic aneurysms and dissections and specific genetic factors (hazard ratio =1.43, P=3.3×10⁻²⁰) (93). This approach provides valuable insights for early diagnosis and targeted intervention in aortic diseases.

In a subsequent study, Pirruccello et al. applied DL models to quantify dynamic aortic properties, such as distensibility and strain, further clarifying their association with cardiovascular diseases. By integrating DL with GWAS, the study identified several genetic loci associated with aortic strain and distensibility, while polygenic risk scores demonstrated strong associations with hypertension, coronary artery disease, CKD, and stroke. These findings showed that dynamic aortic properties provide a more comprehensive basis for cardiovascular disease prediction than traditional aortic diameter measurements (94).

Integrating DL with genetic analysis holds substantial potential, enabling automated imaging and the incorporation of genomic data to advance our understanding of underlying genetic mechanisms and improve risk prediction (95). However, the reliance on European-descent populations and the lack of external validation limit the generalizability and translational potential of these AI-driven models.

Outlook

Advances and trends

AI-driven radiomics has become a transformative tool in vascular imaging, enabling automated feature extraction and individualized risk assessment. Radiomics quantifies imaging features (plaque burden, calcification, and remodeling), enhancing disease detection and prognostic prediction when combined with clinical risk factors (12). Further, radiogenomics reveals the genetic bases of vascular diseases, advancing precision medicine (96).

To illustrate the evolution of research themes in AI-driven vascular imaging, a keyword-based timeline was constructed (Figure 9). The timeline highlights a shift from traditional ML to advanced DL methods. Emerging topics like radiomics, coronary calcium scoring, and radiogenomics are driving advances in precision diagnostics, highlighting the need for innovation and standardized protocols to ensure model reproducibility and clinical translation.

Figure 9 Timeline of keyword clusters based on co-citation analysis. Nodes represent highly cited articles or influential keywords, with their size proportional to their citation frequency. Lines indicate co-citation relationships, illustrating the evolution of research themes over time.

Challenges

Despite advancements, AI-driven vascular imaging faces challenges that limit the development of robust and generalizable applications. The use of limited, biased, and heterogeneous training data impairs the generalizability of AI models (79). Detailing cohort characteristics as per the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) guidelines is essential to ensure model robustness (97). Inconsistent imaging protocols and poor adherence to the Imaging Biomarker Standardisation Initiative (IBSI) standards reduce model reproducibility and impede comparisons (98). Further, as many studies use varying methods, comparisons are difficult. Thus, following IBSI guidelines is essential to improve model reproducibility and reliability. Further, a lack of robust external validation remains a major barrier to the clinical application of such models (79). Many studies lack the independent cohorts required by TRIPOD, resulting in inconsistent performance and highlighting the need for multicenter collaborations. Finally, the “black-box” nature of DL models erodes clinical trust, limiting model adoption. Developing transparent, interpretable architectures is thus essential to ensure the clinical utility of these models (99).

Conclusions

In conclusion, AI is transforming vascular imaging and radiogenomics, paving the way for precision medicine. Future research should prioritize multicenter collaborations and the use of diverse, externally validated datasets to improve model generalizability. Standardizing imaging protocols and feature extraction (as per the IBSI guidelines) is essential to ensure comparability and reproducibility. Advancing explainable AI will increase model transparency and trust. By integrating established frameworks such as TRIPOD (97), the Checklist for Artificial Intelligence in Medical Imaging (CLAIM) (100), and the CheckList for EvaluAtion of Radiomics research (CLEAR) (101), future studies can help to establish guidelines and accelerate the adoption of AI in personalized medicine.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-531/rc

Peer Review File: Available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-531/prf

Funding: This work was supported by Natural Science Foundation of Hubei Province (No. 2023BCB018 to F.H.), National Natural Science Foundation of China (No. 81974089 to F.H.), and the Key Research and Development Program of Hubei Province (No. 2022BCA001).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://cdt.amegroups.com/article/view/10.21037/cdt-2025-531/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.

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. Hurford R, Wolters FJ, Li L, et al. Prevalence, predictors, and prognosis of symptomatic intracranial stenosis in patients with transient ischaemic attack or minor stroke: a population-based cohort study. Lancet Neurol 2020;19:413-21. [Crossref] [PubMed]
2. Cheng K, Lin A, Yuvaraj J, et al. Cardiac Computed Tomography Radiomics for the Non-Invasive Assessment of Coronary Inflammation. Cells 2021;10:879. [Crossref] [PubMed]
3. Blanco PJ, Ziemer PGP, Bulant CA, et al. Fully automated lumen and vessel contour segmentation in intravascular ultrasound datasets. Med Image Anal 2022;75:102262. [Crossref] [PubMed]
4. Chung H, Cha E, Sunwoo L, et al. Two-stage deep learning for accelerated 3D time-of-flight MRA without matched training data. Med Image Anal 2021;71:102047. [Crossref] [PubMed]
5. Kothari G, Woon B, Patrick CJ, et al. The impact of inter-observer variation in delineation on robustness of radiomics features in non-small cell lung cancer. Sci Rep 2022;12:12822. [Crossref] [PubMed]
6. Nature Biomedical Engineering. Auspicious machine learning. Nat Biomed Eng 2017;1:0036.
7. Majumder S, Katz S, Kontos D, et al. State of the art: radiomics and radiomics-related artificial intelligence on the road to clinical translation. BJR Open 2024;6:tzad004. [Crossref] [PubMed]
8. A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence [Internet]. [cited 2024 Oct 10]. Available online: http://jmc.stanford.edu/articles/dartmouth.html
9. Nilsson NJ. The Quest for Artificial Intelligence [Internet]. 1st edn. Cambridge University Press; 2009 [cited 2024 Oct 10]. Available online: https://www.cambridge.org/core/product/identifier/9780511819346/type/book
10. Campbell M, Hoane AJ, Hsu F. Deep Blue. Artificial Intelligence 2002;134:57-83. [Crossref]
11. LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521:436-44. [Crossref] [PubMed]
12. Mastrodicasa D, van Assen M, Huisman M, et al. Use of AI in Cardiac CT and MRI: A Scientific Statement from the ESCR, EuSoMII, NASCI, SCCT, SCMR, SIIM, and RSNA. Radiology [Internet]. 2025 Jan 28 [cited 2025 Apr 19]. Available online: https://pubs.rsna.org/doi/10.1148/radiol.240516
13. Greener JG, Kandathil SM, Moffat L, et al. A guide to machine learning for biologists. Nat Rev Mol Cell Biol 2022;23:40-55. [Crossref] [PubMed]
14. Gu J, Wang Z, Kuen J, et al. Recent advances in convolutional neural networks. Pattern Recognition 2018;77:354-77. [Crossref]
15. Mamo AA, Gebresilassie BG, Mukherjee A, et al. Advancing Medical Imaging Through Generative Adversarial Networks: A Comprehensive Review and Future Prospects. Cogn Comput 2024;16:2131-53. [Crossref]
16. Zhou S, Nie D, Adeli E, et al. High-Resolution Encoder-Decoder Networks for Low-Contrast Medical Image Segmentation. IEEE Trans Image Process 2019; Epub ahead of print. [Crossref] [PubMed]
17. Gago L, Vila MDM, Grau M, et al. An end-to-end framework for intima media measurement and atherosclerotic plaque detection in the carotid artery. Comput Methods Programs Biomed 2022;223:106954. [Crossref] [PubMed]
18. Bajaj R, Huang X, Kilic Y, et al. Advanced deep learning methodology for accurate, real-time segmentation of high-resolution intravascular ultrasound images. Int J Cardiol 2021;339:185-91. [Crossref] [PubMed]
19. Araki T, Ikeda N, Shukla D, et al. PCA-based polling strategy in machine learning framework for coronary artery disease risk assessment in intravascular ultrasound: A link between carotid and coronary grayscale plaque morphology. Comput Methods Programs Biomed 2016;128:137-58. [Crossref] [PubMed]
20. Luan S, Yu X, Lei S, et al. Deep learning for fast super-resolution ultrasound microvessel imaging. Phys Med Biol 2023; [Crossref] [PubMed]
21. Eun DI, Jang R, Ha WS, et al. Deep-learning-based image quality enhancement of compressed sensing magnetic resonance imaging of vessel wall: comparison of self-supervised and unsupervised approaches. Sci Rep 2020;10:13950. [Crossref] [PubMed]
22. He D, Zhou J, Shang X, et al. De-Noising of Photoacoustic Microscopy Images by Attentive Generative Adversarial Network. IEEE Trans Med Imaging 2023;42:1349-62. [Crossref] [PubMed]
23. Mori R, Kassai Y, Masuda A, et al. Ultrashort echo time time-spatial labeling inversion pulse magnetic resonance angiography with denoising deep learning reconstruction for the assessment of abdominal visceral arteries. J Magn Reson Imaging 2021;53:1926-37. [Crossref] [PubMed]
24. Shou Q, Zhao C, Shao X, et al. Transformer-based deep learning denoising of single and multi-delay 3D arterial spin labeling. Magn Reson Med 2024;91:803-18. [Crossref] [PubMed]
25. Rutkowski DR, Roldán-Alzate A, Johnson KM. Enhancement of cerebrovascular 4D flow MRI velocity fields using machine learning and computational fluid dynamics simulation data. Sci Rep 2021;11:10240. [Crossref] [PubMed]
26. Zhou Z, Wang Y, Guo Y, et al. Image Quality Improvement of Hand-Held Ultrasound Devices With a Two-Stage Generative Adversarial Network. IEEE Trans Biomed Eng 2020;67:298-311. [Crossref] [PubMed]
27. Koktzoglou I, Huang R, Ong AL, et al. Feasibility of a sub-3-minute imaging strategy for ungated quiescent interval slice-selective MRA of the extracranial carotid arteries using radial k-space sampling and deep learning-based image processing. Magn Reson Med 2020;84:825-37. [Crossref] [PubMed]
28. Lok UW, Huang C, Gong P, et al. Fast super-resolution ultrasound microvessel imaging using spatiotemporal data with deep fully convolutional neural network. Phys Med Biol 2021; [Crossref] [PubMed]
29. Qi H, Fuin N, Cruz G, et al. Non-Rigid Respiratory Motion Estimation of Whole-Heart Coronary MR Images Using Unsupervised Deep Learning. IEEE Trans Med Imaging 2021;40:444-54. [Crossref] [PubMed]
30. Wood G, Pedersen AU, Kunze KP, et al. Automated detection of cardiac rest period for trigger delay calculation for image-based navigator coronary magnetic resonance angiography. J Cardiovasc Magn Reson 2023;25:52. [Crossref] [PubMed]
31. Küstner T, Munoz C, Psenicny A, et al. Deep-learning based super-resolution for 3D isotropic coronary MR angiography in less than a minute. Magn Reson Med 2021;86:2837-52. [Crossref] [PubMed]
32. Park JH, Lee SJ. Ultrasound Deep Learning for Wall Segmentation and Near-Wall Blood Flow Measurement. IEEE Trans Ultrason Ferroelectr Freq Control 2020;67:2022-32. [Crossref] [PubMed]
33. Yoon YH, Khan S, Huh J, et al. Efficient B-Mode Ultrasound Image Reconstruction From Sub-Sampled RF Data Using Deep Learning. IEEE Trans Med Imaging 2019;38:325-36. [Crossref] [PubMed]
34. Wu P, Qiao Y, Chu M, et al. Reciprocal assistance of intravascular imaging in three-dimensional stent reconstruction: Using cross-modal translation based on disentanglement representation. Comput Med Imaging Graph 2023;104:102166. [Crossref] [PubMed]
35. Jun Y, Eo T, Shin H, et al. Parallel imaging in time-of-flight magnetic resonance angiography using deep multistream convolutional neural networks. Magn Reson Med 2019;81:3840-53. [Crossref] [PubMed]
36. Materka A, Jurek J. Using Deep Learning and B-Splines to Model Blood Vessel Lumen from 3D Images. Sensors (Basel) 2024;24:846. [Crossref] [PubMed]
37. Nazir N, Sarwar A, Saini BS. Recent developments in denoising medical images using deep learning: An overview of models, techniques, and challenges. Micron 2024;180:103615. [Crossref] [PubMed]
38. Mergen V, Sartoretti T, Baer-Beck M, et al. Ultra-High-Resolution Coronary CT Angiography With Photon-Counting Detector CT: Feasibility and Image Characterization. Invest Radiol 2022;57:780-8. [Crossref] [PubMed]
39. Lloyd DFA, Pushparajah K, Simpson JM, et al. Three-dimensional visualisation of the fetal heart using prenatal MRI with motion-corrected slice-volume registration: a prospective, single-centre cohort study. Lancet 2019;393:1619-27. [Crossref] [PubMed]
40. Abdolmanafi A, Duong L, Dahdah N, et al. Intra-Slice Motion Correction of Intravascular OCT Images Using Deep Features. IEEE J Biomed Health Inform 2019;23:931-41. [Crossref] [PubMed]
41. Wu X, Tang L, Li W, et al. Feasibility of accelerated non-contrast-enhanced whole-heart bSSFP coronary MR angiography by deep learning-constrained compressed sensing. Eur Radiol 2023;33:8180-90. [Crossref] [PubMed]
42. Golledge J. Update on the pathophysiology and medical treatment of peripheral artery disease. Nat Rev Cardiol 2022;19:456-74. [Crossref] [PubMed]
43. Shi X, Gao J, Lv Q, et al. Calcification in Atherosclerotic Plaque Vulnerability: Friend or Foe? Front Physiol 2020;11:56. [Crossref] [PubMed]
44. Grubic N, Colledanchise KN, Liblik K, et al. The Role of Carotid and Femoral Plaque Burden in the Diagnosis of Coronary Artery Disease. Curr Cardiol Rep 2020;22:121. [Crossref] [PubMed]
45. He Z, Luo J, Lv M, et al. Characteristics and evaluation of atherosclerotic plaques: an overview of state-of-the-art techniques. Front Neurol 2023;14:1159288. [Crossref] [PubMed]
46. Peper ES, van Ooij P, Jung B, et al. Advances in machine learning applications for cardiovascular 4D flow MRI. Front Cardiovasc Med 2022;9:1052068. [Crossref] [PubMed]
47. Stone PH, Libby P, Boden WE. Fundamental Pathobiology of Coronary Atherosclerosis and Clinical Implications for Chronic Ischemic Heart Disease Management-The Plaque Hypothesis: A Narrative Review. JAMA Cardiol 2023;8:192-201. [Crossref] [PubMed]
48. Li L, Hu Z, Huang Y, et al. Automatic multi-plaque tracking and segmentation in ultrasonic videos. Med Image Anal 2021;74:102201. [Crossref] [PubMed]
49. Lin A, Kolossváry M, Cadet S, et al. Radiomics-Based Precision Phenotyping Identifies Unstable Coronary Plaques From Computed Tomography Angiography. JACC Cardiovasc Imaging 2022;15:859-71. [Crossref] [PubMed]
50. Zhang R, Zhang Q, Ji A, et al. Identification of high-risk carotid plaque with MRI-based radiomics and machine learning. Eur Radiol 2021;31:3116-26. [Crossref] [PubMed]
51. Choi AD, Marques H, Kumar V, et al. CT Evaluation by Artificial Intelligence for Atherosclerosis, Stenosis and Vascular Morphology (CLARIFY): A Multi-center, international study. J Cardiovasc Comput Tomogr 2021;15:470-6. [Crossref] [PubMed]
52. Cho GW, Ghanem AK, Quesada CG, et al. Quantitative plaque analysis with A.I.-augmented CCTA in end-stage renal disease and complex CAD. Clin Imaging 2022;89:155-61. [Crossref] [PubMed]
53. Cho GW, Anderson L, Quesada CG, et al. Serial analysis of coronary artery disease progression by artificial intelligence assisted coronary computed tomography angiography: early clinical experience. BMC Cardiovasc Disord 2022;22:506. [Crossref] [PubMed]
54. Greenland P, Blaha MJ, Budoff MJ, et al. Coronary Calcium Score and Cardiovascular Risk. J Am Coll Cardiol 2018;72:434-47. [Crossref] [PubMed]
55. Gupta A, Bera K, Kikano E, et al. Coronary Artery Calcium Scoring: Current Status and Future Directions. Radiographics 2022;42:947-67. [Crossref] [PubMed]
56. Zhang N, Yang G, Zhang W, et al. Fully automatic framework for comprehensive coronary artery calcium scores analysis on non-contrast cardiac-gated CT scan: Total and vessel-specific quantifications. Eur J Radiol 2021;134:109420. [Crossref] [PubMed]
57. Sartoretti E, Gennari AG, Maurer A, et al. Opportunistic deep learning powered calcium scoring in oncologic patients with very high coronary artery calcium (≥ 1000) undergoing 18F-FDG PET/CT. Sci Rep 2022;12:19191. [Crossref] [PubMed]
58. Feuchtner GM, Lacaita PG, Bax JJ, et al. AI-Quantitative CT Coronary Plaque Features Associate With a Higher Relative Risk in Women: CONFIRM2 Registry. Circ Cardiovasc Imaging 2025;18:e018235. [Crossref] [PubMed]
59. Griffin WF, Choi AD, Riess JS, et al. AI Evaluation of Stenosis on Coronary CTA, Comparison With Quantitative Coronary Angiography and Fractional Flow Reserve: A CREDENCE Trial Substudy. JACC Cardiovasc Imaging 2023;16:193-205. [Crossref] [PubMed]
60. Yuan N, Kwan AC, Duffy G, et al. Prediction of Coronary Artery Calcium Using Deep Learning of Echocardiograms. J Am Soc Echocardiogr 2023;36:474-481.e3. [Crossref] [PubMed]
61. Dobrolinska MM, Lazarenko SV, van der Zant FM, et al. Performance of visual, manual, and automatic coronary calcium scoring of cardiac (13)N-ammonia PET/low dose CT. J Nucl Cardiol 2023;30:239-50. [Crossref] [PubMed]
62. Johri AM, Singh KV, Mantella LE, et al. Deep learning artificial intelligence framework for multiclass coronary artery disease prediction using combination of conventional risk factors, carotid ultrasound, and intraplaque neovascularization. Comput Biol Med 2022;150:106018. [Crossref] [PubMed]
63. Dey D, Slomka PJ, Leeson P, et al. Artificial Intelligence in Cardiovascular Imaging: JACC State-of-the-Art Review. J Am Coll Cardiol 2019;73:1317-35. [Crossref] [PubMed]
64. Panayides AS, Amini A, Filipovic ND, et al. AI in Medical Imaging Informatics: Current Challenges and Future Directions. IEEE J Biomed Health Inform 2020;24:1837-57. [Crossref] [PubMed]
65. Reference Values for Arterial Stiffness' Collaboration. Determinants of pulse wave velocity in healthy people and in the presence of cardiovascular risk factors: 'establishing normal and reference values'. Eur Heart J 2010;31:2338-50. [Crossref] [PubMed]
66. Sahani AK, Srivastava D, Sivaprakasam M, et al. A Machine Learning Pipeline for Measurement of Arterial Stiffness in A-Mode Ultrasound. IEEE Trans Ultrason Ferroelectr Freq Control 2022;69:106-13. [Crossref] [PubMed]
67. Karageorgos GM, Liang P, Mobadersany N, et al. Unsupervised deep learning-based displacement estimation for vascular elasticity imaging applications. Phys Med Biol 2023; [Crossref] [PubMed]
68. Lin A, van Diemen PA, Motwani M, et al. Machine Learning From Quantitative Coronary Computed Tomography Angiography Predicts Fractional Flow Reserve-Defined Ischemia and Impaired Myocardial Blood Flow. Circ Cardiovasc Imaging 2022;15:e014369. [Crossref] [PubMed]
69. Anwer AMOA, Karacan H, Enver L, et al. Machine learning applications for vascular stenosis detection in computed tomography angiography: a systematic review and meta-analysis. Neural Comput & Applic 2024;36:17767-86. [Crossref]
70. Jin X, Li Y, Yan F, et al. Automatic coronary plaque detection, classification, and stenosis grading using deep learning and radiomics on computed tomography angiography images: a multi-center multi-vendor study. Eur Radiol 2022;32:5276-86. [Crossref] [PubMed]
71. Wang X, Yang YQ, Cai S, et al. Deep-learning-based sampling position selection on color Doppler sonography images during renal artery ultrasound scanning. Sci Rep 2024;14:11768. [Crossref] [PubMed]
72. Campinho P, Vilfan A, Vermot J. Blood Flow Forces in Shaping the Vascular System: A Focus on Endothelial Cell Behavior. Front Physiol 2020;11:552. [Crossref] [PubMed]
73. Garrido-Oliver J, Aviles J, Córdova MM, et al. Machine learning for the automatic assessment of aortic rotational flow and wall shear stress from 4D flow cardiac magnetic resonance imaging. Eur Radiol 2022;32:7117-27. [Crossref] [PubMed]
74. He B, Lei J, Lang X, et al. Ultra-fast ultrasound blood flow velocimetry for carotid artery with deep learning. Artif Intell Med 2023;144:102664. [Crossref] [PubMed]
75. Bratt A, Kim J, Pollie M, et al. Machine learning derived segmentation of phase velocity encoded cardiovascular magnetic resonance for fully automated aortic flow quantification. J Cardiovasc Magn Reson 2019;21:1. [Crossref] [PubMed]
76. Cai L, Zhao E, Niu H, et al. A machine learning approach to predict cerebral perfusion status based on internal carotid artery blood flow. Comput Biol Med 2023;164:107264. [Crossref] [PubMed]
77. Obuchowicz R, Strzelecki M, Piórkowski A. Clinical Applications of Artificial Intelligence in Medical Imaging and Image Processing-A Review. Cancers (Basel) 2024;16:1870. [Crossref] [PubMed]
78. Corti A, Lo Iacono F, Ronchetti F, et al. Enhancing cardiovascular risk stratification: Radiomics of coronary plaque and perivascular adipose tissue - Current insights and future perspectives. Trends Cardiovasc Med 2025;35:47-59. [Crossref] [PubMed]
79. Yang Y, Zhang H, Gichoya JW, et al. The limits of fair medical imaging AI in real-world generalization. Nat Med 2024;30:2838-48. [Crossref] [PubMed]
80. Saba L, Maindarkar M, Johri AM, et al. UltraAIGenomics: Artificial Intelligence-Based Cardiovascular Disease Risk Assessment by Fusion of Ultrasound-Based Radiomics and Genomics Features for Preventive, Personalized and Precision Medicine: A Narrative Review. Rev Cardiovasc Med 2024;25:184. [Crossref] [PubMed]
81. Gale MS. Diagnosis: Fundamental Principles and Methods. Cureus 2022;14:e28730. [PubMed]
82. Mikail N, Rossi A, Bengs S, et al. Imaging of heart disease in women: review and case presentation. Eur J Nucl Med Mol Imaging 2022;50:130-59. [Crossref] [PubMed]
83. Chen H, Yan S, Xie M, et al. Application of cascaded GAN based on CT scan in the diagnosis of aortic dissection. Comput Methods Programs Biomed 2022;226:107130. [Crossref] [PubMed]
84. Claux F, Baudouin M, Bogey C, et al. Dense, deep learning-based intracranial aneurysm detection on TOF MRI using two-stage regularized U-Net. J Neuroradiol 2023;50:9-15. [Crossref] [PubMed]
85. Wu J, Xin J, Yang X, et al. Deep morphology aided diagnosis network for segmentation of carotid artery vessel wall and diagnosis of carotid atherosclerosis on black-blood vessel wall MRI. Med Phys 2019;46:5544-61. [Crossref] [PubMed]
86. Bera K, Braman N, Gupta A, et al. Predicting cancer outcomes with radiomics and artificial intelligence in radiology. Nat Rev Clin Oncol 2022;19:132-46. [Crossref] [PubMed]
87. Zhang X, Miao J, Yang J, et al. DWI-Based Radiomics Predicts the Functional Outcome of Endovascular Treatment in Acute Basilar Artery Occlusion. AJNR Am J Neuroradiol 2023;44:536-42. [Crossref] [PubMed]
88. Luo Y, Sun X, Kong X, et al. A DWI-based radiomics-clinical machine learning model to preoperatively predict the futile recanalization after endovascular treatment of acute basilar artery occlusion patients. Eur J Radiol 2023;161:110731. [Crossref] [PubMed]
89. Feng C, Chen R, Dong S, et al. Predicting coronary plaque progression with conventional plaque parameters and radiomics features derived from coronary CT angiography. Eur Radiol 2023;33:8513-20. [Crossref] [PubMed]
90. Lee SE, Hong Y, Hong J, et al. Prediction of the development of new coronary atherosclerotic plaques with radiomics. J Cardiovasc Comput Tomogr 2024;18:274-80. [Crossref] [PubMed]
91. Liu Y, Kong Y, Yan Y, et al. Explore the value of carotid ultrasound radiomics nomogram in predicting ischemic stroke risk in patients with type 2 diabetes mellitus. Front Endocrinol (Lausanne) 2024;15:1357580. [Crossref] [PubMed]
92. Infante T, Cavaliere C, Punzo B, et al. Radiogenomics and Artificial Intelligence Approaches Applied to Cardiac Computed Tomography Angiography and Cardiac Magnetic Resonance for Precision Medicine in Coronary Heart Disease: A Systematic Review. Circ Cardiovasc Imaging 2021;14:1133-46. [Crossref] [PubMed]
93. Pirruccello JP, Chaffin MD, Chou EL, et al. Deep learning enables genetic analysis of the human thoracic aorta. Nat Genet 2022;54:40-51. [Crossref] [PubMed]
94. Pirruccello JP, Rämö JT, Choi SH, et al. The Genetic Determinants of Aortic Distention. J Am Coll Cardiol 2023;81:1320-35. [Crossref] [PubMed]
95. Steinfeldt J, Buergel T, Loock L, et al. Neural network-based integration of polygenic and clinical information: development and validation of a prediction model for 10-year risk of major adverse cardiac events in the UK Biobank cohort. Lancet Digit Health 2022;4:e84-94. [Crossref] [PubMed]
96. Fritzsche MC, Akyüz K, Cano Abadía M, et al. Ethical layering in AI-driven polygenic risk scores-New complexities, new challenges. Front Genet 2023;14:1098439. [Crossref] [PubMed]
97. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): the TRIPOD Statement. [cited 2024 July 17]. Available online: https://core.ac.uk/reader/81583807?utm_source=linkout
98. Zwanenburg A, Vallières M, Abdalah MA, et al. The Image Biomarker Standardization Initiative: Standardized Quantitative Radiomics for High-Throughput Image-based Phenotyping. Radiology 2020;295:328-38. [Crossref] [PubMed]
99. Huang EP, O'Connor JPB, McShane LM, et al. Criteria for the translation of radiomics into clinically useful tests. Nat Rev Clin Oncol 2023;20:69-82. [Crossref] [PubMed]
100. Mongan J, Moy L, Kahn CE Jr. Checklist for Artificial Intelligence in Medical Imaging (CLAIM): A Guide for Authors and Reviewers. Radiol Artif Intell 2020;2:e200029. [Crossref] [PubMed]
101. Kocak B, Baessler B, Bakas S, et al. CheckList for EvaluAtion of Radiomics research (CLEAR): a step-by-step reporting guideline for authors and reviewers endorsed by ESR and EuSoMII. Insights Imaging 2023;14:75. [Crossref] [PubMed]