Artificial Intelligence in Cardiovascular Imaging: Current Applications and New Horizons

Abstract

Artificial intelligence (AI) is transforming cardiovascular imaging (CVI), enhancing accuracy, efficiency, and diagnostic capability across echocardiography (Echo), cardiac computed tomography (CCT), and cardiac magnetic resonance (CMR). In Echo, AI improves image acquisition, segmentation, quantification of chamber function, and detection of wall motion abnormalities, supporting diagnosis and prognosis in various diseases. Automated two-dimensional and three-dimensional (3D) analysis allows rapid, reproducible assessments of ventricular volumes and EF. In valvular heart disease, AI assists in measurement, procedural planning, and integration with 3D printing. CCT benefits from AI at every workflow stage, from image acquisition to disease assessment. AI optimizes scanning protocols, reduces radiation exposure, and enhances coronary artery calcium scoring, plaque analysis, and ischemia evaluation. Algorithms enable rapid segmentation and functional assessment, while ongoing studies support its utility in risk prediction and plaque characterization. In CMR, AI accelerates acquisition, reduces artifacts, and automates segmentation and tissue characterization. Deep learning (DL) models accurately detect fibrosis, scar, and functional parameters, positively influencing prognosis prediction in every cardiac disease. AI-driven tools also streamline report generation, enhance Telemedicine workflow, and guide less experienced users in image acquisition. Despite these advances, challenges remain. Robust and diverse datasets, explainable AI models, regulatory approvals, and ethical considerations are critical for safe and widespread adoption. AI’s “black box” nature hinders clinician trust, making interpretability essential. As these barriers are addressed, AI is expected to become an essential tool in every aspect of CVI, enabling personalized medicine, improving patient care, and optimizing clinical workflows in the coming decades.

INTRODUCTION

Artificial intelligence (AI) is now ubiquitous and is applied in many sectors. Research into the use of AI in health care is one of the most active areas, and AI has already achieved impressive results. AI has been widely applied in the field of cardiovascular imaging (CVI), including echocardiography (Echo), cardiac computed tomography (CCT), and cardiac magnetic resonance (CMR).

In this review, we explore the main clinical applications of AI in CVI and the principal methodological AI approaches that have been developed to solve the related image analysis problems.

GENERAL ASPECTS AND CONSIDERATIONS

The most prominent AI tool currently being used is deep learning, which is highly efficient at extracting spatial and temporal associations from large databases. This review is not intended to address AI technical issues. However, supervised learning is a widely applicable technique in medical imaging, in which both input information (such as data from Echo or CMR) and output information (such as the [EF] of the left ventricle [LV]), considered to be ground truth data, are provided to the machine learning algorithm in the training data in order to learn a mapping from input to output information.

Many different aspects of CVI can be improved by AI: patient selection and referral; identification of patients who can benefit most from imaging and with the use of which modality; improvement of diagnostic workflow and setting of the machines; identification of cardiac structures and detection of anomalies; automated measurements; diagnosis and prognosis of specific pathologies; therapy selection; and planning and guidance of interventional and surgical procedures.

We will highlight the methods, benefits, and limitations of AI in the three main areas of imaging: Echo, CT, and CMR.

ARTIFICAIL INTELLIGENCE AND ECHOCARDIOGRAPHY

In all Echo techniques and modalities, AI may improve imaging quality, guiding scanning, and assisting in segmentation, processing, and analysis. It can also help in view interpretation and classification, in the quantification of left and right chamber volumes and function, and in detecting wall motion abnormalities. Moreover, AI can also help in the diagnosis and prognosis of several cardiovascular diseases and assist in the measurements and simulation of structural procedures.[1,2,3,4,5]

Figure 1 summarizes a scheme of the pipeline and potentiality of AI in Echo from acquisition to analysis, risk assessment, and guidance techniques.

Figure 1.

Pipeline and potentiality of artificial intelligence in echocardiography from acquisition to analysis, risk assessment and guidance techniques

ACQUISITION AND QUANTIFICATION OF LEFT AND RIGHT CHAMBER VOLUMES AND FUNCTION

Nowadays several ultrasound units of different companies adopt AI systems for rapid assessment of echo data, echo volumes, Doppler data and mainly three-dimensional (3D) volumes and function measurements. Machine learning (ML) algorithms have resulted in a 2D or 3D analysis technique that automatically detects LV and left atrium (LA) boundaries and allows fast, accurate, automated measurements of chamber volumes and function. Knackstedt et al. showed that a novel, fully automated software using ML-enabled image analysis provides rapid, reproducible measurements of LV volume and ejection fraction (EF) and longitudinal strain.[6] Tsang et al. first demonstrated that automated simultaneous quantification of LA and LV volumes and LVEF was feasible with minimal 3D software analysis training.[7] Moreover, these 3D TTE automated measurements were comparable to CMR values, and timesaving (approximately 30 s). The method is easy since 3D Echo acquisitions are obtained from the four-chamber apical view in full-volume mode during a breath-hold lasting a few seconds. The program identifies LV end-diastole and end-systole using the ECG and determines global cardiac shape orientation (both LV and LA). Database was then built on the basis of 1000 3D TTE datasets with a wide range of functions and morphologies (i.e., normal, dilated, hypertrophic phenotype).[8] Several subsequent studies further improved the method including a dynamic evaluation that allowed to obtain functional curves of LV and LA.[9] Figure 2 shows an example of 3DTTE acquisition and AI reconstruction of the LV and LA in a normal subject.

Figure 2.

3D-TTE acquisition and artificial intelligence reconstruction of the left ventricle and left atrium in a normal subject

Volume and EF obtained with these AI systems have been validated against CMR and may represent a valid alternative to the traditional measurements for everyday clinical use not only in selected cases but also in unselected population.[10] Ouyang et al. proposed a new video DL algorithm that achieves state-of-the-art assessment of cardiac function. This system surpasses human expert performance in the critical tasks of segmenting the LV, estimating LVEF, and assessing cardiomyopathies.[11]

Similarly, a new AI-based automated 3DEcho software for the RV evaluation has been proposed and validated versus CMR. The RV plays a pivotal role in cardiovascular diseases and 3D Echo has gained acceptance for the evaluation of RV volumes and function.[12] This is extremely relevant because the Echo study of the function and size of the RV has always been of great interest for its diagnostic and prognostic role in many cardiac pathologies, but has been hindered by the extremely complex morphological and functional structure of the RV. A new AI software nowadays allows a comprehensive evaluation of the RV in <20 s, avoiding the cumbersome process,[13] with normal values for 3D RV volumes and functional parameters parsed by sex and age described by Addetia et al.[14] The acquisition is very easy from an adapted four-chamber apical view including the entire RV and datasets were analyzed on board in few seconds. The final reconstruction includes the inflow, apical, and outflow segments of the RV. When RV endocardial contours are considered suboptimal the operator could intervene manually, editing corrections in few seconds. From the 3D data set 2D measurements were also automatically produced including Fractional Area Change, TAPSE, and RV free wall longitudinal strain.[15] These data are validated against CMR and provide major advantages in the majority of pathologies (prognostic value) and specifically are very useful in patients after cardiac surgery, in whom TAPSE has no clinical value, in the selection of cases undergoing cardiac surgery and percutaneous procedures and in LV device assistance.[16]

ANALYSIS OF OTHER FUNCTIONAL ECHO DATA AND DETECTION OF WALL MOTION ABNORMALITIES

The assessment of regional wall motion abnormalities (RWMAs) is paramount for the Echo evaluation of ischemic heart disease. It is key to the identification of acute and chronic myocardial infarction, as well as the differentiation of ischemic from nonischemic causes of cardiomyopathy. Currently, the assessment of RWMA relies on qualitative interpretation of the multiple Echo views. Recent applications of AI in Echo have shown promise in the field of automated image selection and detection of RWMA.[17] AI reaches excellent accuracy in the identification of RWMA, equivalent to that of expert readers. Moreover, these models outperformed the majority of novice readers. As concerns Stress-Echo, AI has the potential to correct the three main limitations of SE, which are the inter-observer variability, the suboptimal capability of quantification, and the difficulty to translate imaging data into clinical information easily accessible to the end-user cardiologist. Despite technical challenges, AI has already shown a clear potential to improve quantification and reduce variability, and at the same time minimize the analysis time and improve workflow efficiency.[17,18,19,20,21,22]

DIAGNOSIS AND PROGNOSIS

A number of promising studies using DL and clustering approaches have been published for the classification and prediction of cardiac diseases. Zhang et al. presented a fully automated Echo interpretation pipeline for disease detection by training the network using three random images per video as an input and providing two prediction outputs (i.e., diseased or normal). The ROC curve performance of their model for the prediction of hypertrophic cardiomyopathy, cardiac amyloidosis, and pulmonary hypertension were 0.93, 0.87, and 0.85, respectively.[17] Ghorbani et al. trained a customized convolutional neural network (CNN) model that includes a dataset of >1.6 million echocardiogram images from 2850 patients to identify local cardiac structures, estimate cardiac function, and predict systemic risk factors.[23]

Very recently, it has been demonstrated that AI models adapted for use with POCUS (point-of-care ultrasound) can reliably identify hypertrophic cardiomyopathy and transthyretin amyloid cardiomyopathy at the point of care.[24] A supervised ML classifier was also used to discriminate between restrictive cardiomyopathy and constrictive pericarditis using clinical and Echo demonstrating an excellent area under the curve (AUC) of 96.2%.[25] By combining ECG and Echo data, Soto et al. demonstrated a high discriminatory ability in distinguishing hypertrophic cardiomyopathy from hypertension with an AUC of 0.91.[26]

A huge number of other studies have been published demonstrating that AI is becoming an important automated diagnostic tool in the field of pulmonary hypertension, cardiac amyloidosis, and valve pathologies (aortic stenosis and mitral regurgitation).[27,28,29,30]

Several AI-based techniques have been developed to improve the entire imaging chain and optimize image acquisition, and image reconstruction, improve segmentation, and provide fast quantification, diagnosis, and prognostication.[31] Furthermore, new software provides not only automatic measurements but also generates reports by determining normal or abnormal findings based on the latest guideline criteria.

VALVE ANALYSIS, SIMULATION, AND USEFULNESS IN PROCEDURES

The application of AI to Echo images in patients with valve heart disease falls into four main categories: image acquisition, view recognition, image segmentation, and disease state identification. There are also programs to improve guide image acquisition and allow automated measurements. Segmentation can be performed of the valve annulus, leaflets/cusps, jets, and Doppler spectral profiles.[32]

The application of AI in interventional cardiology (IC) can be divided into two main branches, virtual and physical. The virtual branch includes informatics, and cognitive computing to control health management systems, and automated clinical decision support systems. The physical branch is best represented by robotic IC procedures. However, as with many emerging technologies, the true promise of AI within IC may be lost if it is not developed correctly.[33]

Accurate measurement of the aortic valve annulus and root is important in the treatment of aortic valve disease. In this regard, AI software that uses 3D Echo to model the aortic annulus has been validated against cardiac CT.[34]

Very recently, an automated pipeline was developed to identify clinically significant tricuspid regurgitation with excellent performance. Sensitivity and specificity in identifying cases with moderate or severe tricuspid regurgitation were very high thanks to apical four-chamber videos with color Doppler across the tricuspid valve.[35]

Combining AI with the latest developments in 3D printing has enabled the manufacturing of patient-specific anatomic replicas, which yields a significant contribution toward precision medicine. These methods have been applied in several fields of procedures such as transcatheter aortic valve replacement, percutaneous MV repair and prosthesis implantation, and percutaneous tricuspid valve repair and replacement.

TRAINING, GUIDING, AND TELEMEDICINE

ML methods may identify structures within an image and accurately label them, such as identifying an image as a parasternal long-axis view rather than an apical long-axis view (including labeling of all cardiac structures). This has a relevant impact not only in teaching and education but also in allowing to perform a POCUS Echo by a less trained sonographer or physician. These methods facilitate medical residents in the use of handheld (or portable) echo by providing real-time feedback on optimal probe position and acquisition of correct views. This is also an important option in telemedicine and second opinion exchange. Mor-Avi et al. recently demonstrated that a new algorithm for real-time guidance of ultrasound imaging of the heart is feasible and in the hands of nurses and medical residents with minimal training, can yield images of similar diagnostic quality to those obtained by expert sonographers.[36] AI guidance tool is by no means intended to replace comprehensive Echo examination by a professional sonographer, but rather to allow novices to obtain with guidance a limited set of images that may aid in an initial evaluation of a patient in a setting where specialized, trained personnel is not available. Naser et al. proposed an automated view classifier CNN that was able to classify cardiac views obtained using TTE and POCUS with high accuracy.[37]

ARTIFICIAL INTELLIGENCE IN CARDIAC COMPUTED TOMOGRAPHY

Multiple potential applications of AI in CCT have been described in recent years. AI may be integrated across the entire CCT workflow-from patient selection and image acquisition to advanced image interpretation and clinical decision-making. Initial studies focusing on image acquisition and reconstruction demonstrated that AI can effectively optimize scanning protocols and enhance reconstruction algorithms, thus improving image quality while reducing radiation exposure.[38]

This review will outline the major contributions of AI and ML in CCT: AI enables optimization of acquisition protocols, reduction of radiation dose, and enhancement of image quality; AI improves the detection of coronary artery disease (CAD) and allows accurate quantification of atherosclerosis; AI facilitates automated segmentation of cardiac structures, supporting precise preprocedural planning for structural interventions.

ARTIFICIAL INTELLIGENCE FOR IMAGE ACQUISITION

AI-based tools can assist radiographers and technologists in consistently acquiring high-quality cardiovascular CT scans by tailoring acquisition parameters (e.g., kVp and mA) based on scout images. In particular, AI can refine image reconstruction techniques, contributing to further dose reduction.[39] Moreover, DL algorithms can perform noise reduction in low-dose scans by learning from datasets acquired at both low and high radiation doses can facilitate de-noising of low-radiation dose images.[40,41] Additionally, AI can support motion correction by preserving anatomical details while estimating coronary motion direction and magnitude.[42]

ARTIFICIAL INTELLIGENCE FOR CORONARY ARTERY CALCIUM SCORING

CAC scoring is a well-established predictor of obstructive CAD, independent of traditional clinical risk factors, and typically requires semi-automatic image segmentation and manual expert input on ECG-gated CT scans. AI-based algorithms have been shown to perform fully automated CAC scoring on noncontrast cardiac CT with excellent agreement with conventional semiautomatic methods.[40] Automated CAC quantification powered by AI could become an essential tool for patient risk stratification, assisting clinicians in defining cardiovascular risk profiles in both acute and chronic CAD settings.

ARTIFICIAL INTELLIGENCE FOR CORONARY ARTERY DISEASE DETECTION AND CHARACTERIZATION

AI algorithms can segment the coronary artery lumen and vessel wall, enabling renewed interest in quantitative CCT analysis AI can be applied to CCT to segment the coronary lumen and vessel wall. These advances have sparked renewed interest in quantitative coronary computed tomography angiography (CCTA) analysis.[43,44,45] Automated segmentation involves several steps, including anatomical identification, branch labeling, lumen centreline extraction, and definition of inner and outer vessel boundaries. This enables accurate quantification of stenoses and assessment of their severity relative to adjacent vessel segments. Coronary plaque – defined as the area between inner and outer contours – can be characterized in terms of volume, distribution, and composition using ML-based tools, typically relying on CT attenuation thresholds.

One of the most significant AI applications in CCT lies in the detection and characterization of CAD. The recent CLARIFY study (CT Evaluation by AI for Atherosclerosis, Stenosis, and Vascular Morphology) demonstrated that AI-based CCT evaluation achieved excellent diagnostic performance in detecting significant coronary stenoses (>50%), with a sensitivity of 80% and specificity of 97% compared with expert readers.[46] Given that CCT is the only noninvasive modality capable of assessing coronary atherosclerosis-a recognized surrogate endpoint for clinical trials-improving reproducibility and reducing analysis time are critical goals. While conventional quantitative plaque analysis may require up to 30 min, AI significantly accelerates this process.[47] The multicenter PLAQUE study recently showed a strong correlation between AI-driven CCTA plaque quantification and intravascular ultrasound findings, supporting the reliability of AI-based plaque assessment.[48] However, it is important to acknowledge that not all studies have shown favorable results. A sub-analysis of the CLARIFY trial evaluating AI for detecting high-risk plaque features (e.g., positive remodeling, low attenuation) reported poor agreement between AI algorithms and expert readers (weighted Kappa values of 0.22, 0.17, and 0.26, respectively).[49] Future directions for AI in CCTA may include predicting plaque vulnerability and acute coronary syndromes, although current evidence remains limited in this area.

ARTIFICIAL INTELLIGENCE FOR FAT QUANTIFICATION

AI has also been applied to the analysis of pericoronary adipose tissue attenuation, an emerging marker of cardiovascular risk.[50,51] Epicardial adipose tissue, a metabolically active fat depot surrounding the heart, can be quantified using DL techniques on cardiac CT, showing potential for improving risk prediction of adverse cardiovascular events.[52]

ARTIFICIAL INTELLIGENCE FOR ISCHEMIA ASSESSMENT

ML techniques can aid in evaluating the functional significance of coronary artery stenoses on CCTA. Fractional flow reserve derived from CCTA (FFR-CT), which relies on computational fluid dynamics, benefits from ML-based segmentation of coronary anatomy. Alternatively, ML models may predict hyperaemic pressure gradients based solely on morphological features extracted from CT images.[53,54] DL algorithms have also been developed to assess the presence of functionally significant CAD directly from rest CCTA images, achieving promising diagnostic accuracy.[55] Moreover, integrating anatomical, plaque, and functional data allows automated calculation of CAD-RADS scores or other disease classification schemes. An alternative approach involves ML-based analysis of the left ventricular myocardium from rest CCTA to identify functionally significant stenoses, as demonstrated by Zreik et al., achieving an area AUC) of 0.74.[56]

Finally, myocardial perfusion imaging using cardiac CT provides both anatomical and functional data from a single test. Although promising, AI applications in CT perfusion imaging remain limited, with initial studies focusing on the ML-based detection of perfusion defects in rest-phase CCTA.[57]

ARTIFICIAL INTELLIGENCE FOR CARDIAC STRUCTURE SEGMENTATION

Beyond coronary analysis, AI algorithms enable automated segmentation of cardiac chambers (atria and ventricles) and vessels on both contrast-enhanced and noncontrast CT scans. This allows accurate measurement of ventricular volumes, ejection fraction, and myocardial mass. These capabilities have clear clinical utility in the preprocedural planning of structural heart interventions, such as transcatheter aortic valve implantation, where AI-driven segmentation supports precise anatomical assessment.[58]

ARTIFICIAL INTELLIGENCE FOR NONCARDIOVASCULAR FINDINGS

CCTA datasets often include extracardiac structures, and incidental noncardiovascular findings are common, with clinically significant findings reported in 12%–38% of cases.[59] AI can assist as a “second reader” by improving the detection and characterization of these findings, thereby optimizing workflow efficiency. Lung nodules, in particular, are a frequent incidental finding in CCTA, and their detection and characterization are well suited to AI/ML applications.[59]

Figure 3 summarizes fields of cardiac CT in which AI applications can impact daily practice.

Figure 3.

Fields of cardiac computed tomography in which artificial intelligence application can impact in daily practice

ARTIFICIAL INTELLIGENCE IN CARDIAC MAGNETIC RESONANCE

CMR imaging is the gold standard for detailed cardiac structural and functional evaluation. Its superior image resolution, ability to characterize tissue properties, multi-planar imaging capabilities, and absence of ionizing radiation distinguish it from other imaging modalities such as Echo, CT, and nuclear imaging. However, despite its advantages, CMR image acquisition and analysis is still time-consuming. AI offers powerful tools to automate and streamline CMR workflows, enhancing efficiency, consistency, and diagnostic accuracy. In this context, AI-driven methods are rapidly transforming the landscape of CVI, addressing challenges across acquisition, segmentation, and interpretation processes.

ARTIFICIAL INTELLIGENCE IN CARDIAC MAGNETIC RESONANCE IMAGE ACQUISITION

One of the earliest stages where AI can impact CMR is in image acquisition. Conventionally, scan protocol selection and patient positioning are operator dependent, introducing variability and potential errors. AI-driven tools automate these processes, offering standardized prescriptions of cardiac planes and sequence selection based on clinical indications.[60] Vendor-specific AI-based solutions are now integrated into modern MRI scanners, ensuring consistent patient positioning and optimal imaging planes, significantly reducing operator dependency and acquisition times.[61,62] These advances ultimately improve patient comfort and diagnostic yield.

IMAGE RECONSTRUCTION AND ARTIFACT CORRECTION

Long acquisition time in CMR often leads to motion artifacts and patient discomfort. Conventional methods like compressed sensing and parallel imaging have mitigated these challenges to some extent,[63] but AI has introduced a paradigm shift. DL-based models such as 4D CINENet[64] enable accelerated image reconstruction from undersampled data, dramatically reducing reconstruction times while preserving or even enhancing image quality. A study from Eyre et al. showed that a deep learning-accelerated 2D cine CMR method reduces acquisition time by 37% without compromising volumetric measurements or image quality, offering efficient, high-quality imaging suitable for routine clinical use.[65] Moreover, AI techniques facilitate motion artifact detection and correction, using networks like Long-term Recurrent Convolutional Networks to achieve high-fidelity image restoration even in patients unable to perform adequate breath-holding or with arrhythmias.[66]

IMAGE SEGMENTATION

Segmentation of cardiovascular structures is crucial for quantifying parameters such as ventricular volumes, myocardial mass, and scar burden. Manual segmentation is time-consuming and susceptible to intra- and interobserver variability. DL models, particularly U-Net architectures, and novel approaches combining CNNs and Recurrent Neural Networks have further improved temporal consistency across image sequences, enabling reliable and automated assessments that are critical for clinical decision-making, achieving near-human or superior performance in segmenting atria[67] and ventricles.[68]

ARTIFICIAL INTELLIGENCE IN CARDIAC MAGNETIC RESONANCE IMAGE ANALYSIS

Quantitative myocardial mapping techniques, such as T1 and T2 mapping, are essential for detecting myocardial fibrosis, edema, and infiltrative diseases. DL models now automate the estimation of T1 and T2 values from limited input data, achieving results comparable to gold-standard methods such as MOLLI and T2-prepared bSSFP sequences. Techniques, such as DeepBLESS[69] and MyoMapNet,[70] showed high accuracy and rapid processing capabilities of AI in myocardial tissue characterization, thus holding great promise for expanding access to quantitative CMR without significantly extending scan times.

Late gadolinium enhancement (LGE) imaging remains the gold standard for myocardial scar detection, but it is contrast-dependent and time-intensive. Recent AI innovations allow for “virtual” native enhancement using noncontrast images, enabling scar detection comparable to traditional LGE sequences.[71] Moreover, CNNs and both 2D and 3D DL models provide accurate quantification of scar burden, a critical parameter for risk stratification in both ischemic and nonischemic cardiomyopathies.[72] Automated flow quantification using phase-contrast CMR, enhanced by DL models, also improves the assessment of valvular and shunt-related anomalies with excellent reproducibility. In a study by Bratt et al., a ML model was designed to track aortic valve borders based on neural network approaches, and then compared to manual and commercially-available automated segmentation in 190 patients prospectively undergoing CMR. ML segmentation was uniformly successful, extremely quick (1.2 min for the entire dataset compared to 12.5 h for entire dataset with manual segmentation) and greatly correlated with manual segmentation (r = 0.99).[73]

ARTIFICIAL INTELLIGENCE IN DIFFERENT CLINICAL SCENARIOS

AI is increasingly applied in clinical scenarios for diagnosing a range of cardiac pathologies. In hypertrophic cardiomyopathy (HCM), Sahota et al. demonstrated that AI models had the ability to assess LV outflow tract obstruction by analyzing 14 automatically derived landmarks in a retrospective review of 1905 subjects using three-chamber CMR images.[74] Moreover, in the study by Baeßler et al., the texture feature analysis was assessed to identify myocardial tissue changes in HCM in noncontrast T1-weighted CMR images. The authors showed that the model containing a grey level nonuniformity (GLevNonU) texture feature had a sensitivity and specificity of 91% and 93%, respectively, for distinguishing HCM from the control, with an optimal cutoff value for GLevNonU of 46.[75] Similarly, in dilated cardiomyopathy (DCM) ML-based analysis was used to detect patients with the highest chance to positively remodel after medical treatment. The authors evaluated regional LV contractile injury patterns by calculating longitudinal, circumferential, and radial strain over 18 LV sub-regions, building ML models to identify responder subjects in a cohort of 178 test subjects (140 normal subjects and 38 DCM patients). The DNN model predicted response to medical therapy with an AUC of 0.94.[76] Interestingly, in cardiac amyloidosis, Martini et al. showed that DL algorithms built with three networks assessing short axis, two-chamber, and four-chamber LGE images showed similar good diagnostic accuracy compared to ML algorithm considering all manually extracted features (LV volumes, mass and function, LGE pattern, early blood-pool darkening, pericardial and pleural effusion, etc.), reproducing examination reading by an experienced operator (AUC of 0.982 and 0.952, respectively).[77]

ARTIFICIAL INTELLIGENCE IN CARDIAC MAGNETIC RESONANCE AS A PREDICTIVE TOOL

AI-based predictive modeling using CMR-derived features has shown considerable potential in prognosticating clinical outcomes such as heart failure hospitalization, arrhythmia risk, and all-cause mortality. An analysis from the CarDiac MagnEtic Resonance for Primary Prevention Implantable CardioVerter DebrillAtor ThErapy (DERIVATE) registry showed that postinfarct scar with LGE CMR fully automatically quantified by a Ternaus network in a cohort of 761 patients with previous myocardial infarction conferred incremental value over the guideline criterion for the association with arrhythmic events.[78] In the context of stress CMR, Pezel et al. showed that ML predicts 10-year all-cause mortality in patients with suspected or known CAD better than existing clinical or CMR scores. In a retrospective cohort of 31,752 patients ML score (built with automated feature selection by random survival forest, model building with a multiple fractional polynomial algorithm, and five repetitions of 10-fold stratified cross-validation) exhibited a higher AUC compared with Clinical and Stress Cardiac Magnetic Resonance score, European Systematic Coronary Risk Estimation score, QRISK3 score, Framingham Risk Score, and stress CMR data alone for prediction of 10-year all-cause mortality.[79] In a cohort of 758 HCM patients, a ML framework integrating 14 CMR imaging and 23 clinical variables achieved an AUC for the prediction of events (composite of arrhythmic events, SCD, heart failure, and atrial fibrillation-related stroke) of 0.830, outperforming the classic HCM Risk-SCD model with significant improvement of 22.7% in the AUC.[80] Interestingly, in pulmonary hypertension, AI models incorporating 3D RV motion data provide superior survival predictions compared to conventional measures.[81]

PATIENT REPORT GENERATION AND COMMUNICATION

AI is also revolutionizing reporting by automatically generating structured radiology reports based on image analysis, toward a patient-centered communication, simplifying complex findings into understandable language for nonspecialists, thus empowering patients and promoting shared decision-making. Models like KERP (Knowledge-driven Encode, Retrieve, Paraphrase) extract key findings and rephrase them into coherent, clinically relevant reports, improving reporting consistency and reducing physician workload.[82]

Figure 4 represents a summary of the current and potential application of AI-based tools in the CMR field.

Figure 4.

Current and potential application of artificial intelligence-based tools in cardiac magnetic resonance field

LIMITATIONS

Despite remarkable progress, several challenges hinder the widespread clinical adoption of AI in Echo, CCT, and CMR. A primary obstacle is the need for large, diverse, and well-annotated datasets to ensure robust, generalizable models across populations, and imaging platforms. Equally important is regulatory approval and rigorous clinical validation to guarantee safety and efficacy in routine practice. Ethical issues, including data security, patient privacy, and algorithmic accountability, must also be addressed. The evolving regulatory landscape, such as the FDA’s guidance on Software as a Medical Device (SaMD), requires further harmonization to support safe AI deployment. Another major concern is the interpretability of AI algorithms, particularly “black box” models whose decisions lack transparency. Enhancing explainability will be crucial to foster clinician trust and enable meaningful integration of AI-based diagnostic tools into clinical workflows. Overall, overcoming these technical, ethical, and regulatory barriers is essential for the successful integration of AI into CVI.

CONCLUSIONS AND FUTURE CHALLENGES

AI is poised to revolutionize CVI by significantly improving the speed, accuracy, and reproducibility of acquisition, analysis, and reporting workflows [Figure 5]. While commercially available AI tools already offer valuable clinical support, broader acceptance will depend on overcoming challenges related to model interpretability, regulatory approval, data diversity, and medicolegal responsibility. As these hurdles are progressively addressed, the next decade will likely witness the widespread adoption of AI-enhanced Echo, CCT, and CMR, ultimately improving patient outcomes and healthcare efficiency on a global scale.

Figure 5.

Summary of main current application of artificial intelligence in cardiovascular imaging

Conflicts of interest

There are no conflicts of interest.

Funding Statement

Nil.