Radiomics in Cardiovascular Imaging: Current Role and Future Perspectives

Radiomics represents a promising image analysis‏ technique that aims to improve the diagnosis, the‏ characterisation, and the prognosis of diseases by‏ extracting objective quantitative features that may‏ be missed by human eye (Sollini et al. 2019). While‏ mainly developed through oncologic research to obtain‏ information on the characteristics of tumours (Gillies‏ et al. 2016), there is an increasing interest in the use‏ of radiomics for cardiac purposes (Ashrafinia et al.‏ 2021). As cardiovascular diseases represent the main‏ cause of morbidity and mortality worldwide (Virani et‏ al. 2020), there is an ever increasing clinical request for‏ non-invasive diagnostic approaches (Selvanayagam‏ 2016). Therefore, radiomics biomarkers detected by data‏ extraction from cardiac computed tomography (CT) and‏ cardiac magnetic resonance imaging (CMR) may be‏ a valuable tool to assess several cardiac pathologies,‏ such as atherosclerotic coronary artery disease (CAD),‏ myocardial viability, and cardiomyopathies (Kumar et al.‏ 2012; Raisi-Estabragh et al. 2020). Recently, machine‏ learning (ML) and deep learning (DL) algorithms have‏ provided even more options, allowing to better evaluate‏ the characteristics of cardiac disease (Langs et al. 2018).‏

Radiomics in Cardiac CT‏

Cardiac CT angiography (CCTA) has gained a pivotal‏ role in assessing CAD and plays a critical part in‏ evaluating cardiac structures (Hoffmann et al. 2012).‏ CCTA can noninvasively visualise coronary arteries and‏ plaque morphology, representing an invaluable tool for‏ risk stratification, and guide treatment plans in patients‏ with CAD (Douglas et al. 2015).‏

Recently, some exploratory papers have evaluated‏ the feasibility and diagnostic performance of cardiac‏ CT radiomics analysis (Kolossváry et al. 2019; Mannil‏ et al. 2019).‏

In the last few years, some authors developed a‏ radiomics-based ML model that proved to be superior‏ to conventional evaluation of CCTA in the assessment‏ of advanced atheromatous plaques. In particular, the‏ model performed better than radiologists in measuring‏ low attenuation areas and average Hounsfield units‏ of the plaque resulting in a more accurate evaluation‏ of high-risk atherosclerotic lesions, facilitating risk‏ stratification of patients (Kolossváry et al. 2019).‏

Peri-coronary adipose tissue inflammation is another‏ critical element in the development of atherosclerotic‏ plaque, progression, and rupture. Radiomics features‏ extracted from cardiac CT images have demonstrated‏ potential in evaluating the association between‏ atherosclerotic plaques and perivascular adipose tissue‏ inflammation, fibrosis, and vascularity, more precisely‏ than mean attenuation alone (Oikonomou et al. 2019).‏ These results highlighted that the texture phenotype of‏ adipose tissue might provide a non-invasive approach‏ for identifying microvascular adipose tissue remodelling‏ (Antonopoulos et al. 2017; Oikonomou et al. 2019). In‏ some authors’ opinion, this may represent a gamechanger‏ to distinguish patients with acute myocardial‏ infarction from those with stable CAD or to predict‏ patients with a high risk of major adverse cardiac events‏ (MACE) (Lin et al. 2020; Oikonomou et al. 2018).‏

Myocardial tissue characterisation has always been‏ a prerogative of CMR, but with recent technological‏ improvements, even CT scanners can play their part.‏ With the introduction of texture analysis in 2016, CT‏ imaging could discern between healthy and scarred‏ myocardium (Antunes et al. 2016). Soon afterwards, the‏ advent of radiomics analysis increased the capability‏ of cardiac CT in differentiating healthy myocardial‏ tissue from infarcted myocardium. Hinzpeter et al.‏ found that cardiac CT texture analysis was helpful in‏ determining healthy and infarcted myocardial tissue‏ with good reproducibility and accuracy (Hinzpeter et al.‏ 2017). Mannil et al. also demonstrated the capability of‏ radiomics and ML in detecting myocardial infarction on‏ non-contrast CT images acquired for calcium scoring‏ (Mannil et al. 2018).‏

Ventricular arrhythmias (VA) represent an essential‏ prognostic factor in patients with cardiovascular‏ diseases. Researchers have also explored the capability‏ of radiomic features to predict recurrent VA in patients‏ with different remodelling patterns sustained by various‏ cardiomyopathies, such as patients with high arrhythmic‏ risk for left ventricular hypertrophy (Esposito et al. 2018;‏ Kay et al. 2020).‏

Cardiac CT is also essential for the differential‏ diagnosis of cardiac masses in order to establish‏ optimal treatment strategies. However, differentiation is‏ challenging due to the nonspecific clinical and imaging‏ appearances of many cardiac masses (Poterucha et‏ al. 2019). Nam et al. explored the role of CT radiomic‏ features to differentiate pannus from thrombus and‏ vegetation, showing that the algorithms were superior‏ to radiologists in identifying pannus from non-pannus‏ (Nam et al. 2019). Chun et al. compared the capability‏ of radiomics and CT attenuation values in differentiating‏ left atrial appendage thrombus from circulatory stasis‏ and found that the addition of radiomics features‏ represented an added value to help radiologists to‏ identify thrombus in a single early-phase scan (Chun et‏ al. 2021).‏

Despite these exciting applications, they represent‏ only preliminary explorations.‏

These findings, do however, indicate that cardiac‏ CT radiomics may become the next tool to detect‏ image biomarkers more precisely, facilitating improved‏ identification of vulnerable patients.

‏

Radiomics in Cardiac MRI‏

CMR is pivotal in qualitatively and quantitively assessing‏ cardiac structure and function. However, quantitative‏ measures are limited by technical factors and poor‏ discriminatory power due to the overlap of similar‏ appearances of different pathologies. This sometimes‏ makes it challenging to distinguish among similar‏ morphological patterns, such as hypertensive heart‏ disease, hypertrophic cardiomyopathy (HCM) or athletic‏ cardiac remodelling, whose distinctions are critical‏ to guide the clinical assessment, management, and‏ therapy of these patients. Furthermore, CMR likely plays‏ a fundamental role in predicting prognosis in many‏ different clinical settings but its ability is still limited‏ (Moss et al. 2002; Stecker et al. 2006). CMR-based‏ radiomics is emerging as a valid option to help defining‏ image phenotypes and improve diagnosis, prognosis‏ and treatment selection.‏ 

‏Several papers have recently explored the feasibility‏ analysis with CMR.

‏

Some authors have evaluated the ability of CMR‏ radiomics analysis applied to non-contrast cine images‏ to accurately differentiate between myocardial disease‏ states and healthy, suggesting that radiomics features‏ may be capable of highlighting myocardium alteration‏ at a tissue level (Baessler et al. 2018). In particular,‏ a recent study evaluated the ability of radiomics to‏ identify distortions in myocardial architecture that are not‏ detectable by the human eye. These signatures could‏ discriminate accurately between the hearts of individuals‏ with hypertension (morphologically normal at CMR) and‏ those who are normotensive (Cetin et al. 2020).‏

Furthermore, Baessler et al. demonstrated that‏ radiomic texture analysis applied to T1 and T2 maps‏ was superior to mean T1, mean T2, and Lake Louise‏ diagnostic criteria in discriminating infarct-like acute‏ myocarditis (Baessler et al. 2019). The same group also‏ demonstrated the possibility of accurately discerning‏ patients with myocardial infarction from healthy controls‏ through radiomics and texture analysis on CMR cine‏ images without using an intravenous contrast agent,‏ which could be a fundamental achievement for CMR‏ imaging in terms of time-saving and safety (Baessler et‏ al. 2018).‏

In addition, a recently published article stated that‏ radiomics analysis of late gadolinium enhanced‏ (LGE) CMR images is capable of distinguishing acute‏ myocardial infarction from chronic myocardial infarction‏ (Larroza et al. 2017).‏

It is important to remember that the assessment of‏ myocardial infarction and myocardial viability are two of‏ the most frequent requests for which clinicians refer to‏ CMR, and these new radiomic tools could be invaluable‏ in offering better clinical support.‏

As to prognosis and prediction of clinical outcomes, a‏ few studies have been recently published. In a study on‏ patients with chronic myocardial infarction, Kotu et his‏ group demonstrated that radiomics features extracted‏ from LGE scar were superior to scar size and location‏ in determining the risk of dangerous arrhythmias (Kotu‏ et al. 2015). Similarly, Amano et al. showed that different‏ textural features extracted from LGE images could be‏ useful to predict VA in HCM patients (Amano et al. 2018).‏

In summary then, CMR radiomics has the potential‏ to improve myocardial disease classification and‏ prognosis. However, the literature about CMR radiomics‏ is still limited, and further efforts are needed to confirm‏ these preliminary results and to facilitate the radiomics‏ transition from academic to clinical settings.

‏

Conclusion‏

Through specific, quantitative insights provided at a‏ microstructural level, radiomics can favour a better‏ understanding of the physiologic mechanisms of cardiac‏ disease that can be of great value in developing tailored‏ cardiovascular medicine therapy. And thus, in future,‏ be gradually introduced into the clinical physicians’‏ workflow.‏

However, even if most of the recently published papers‏ have shown promising results applying radiomics to‏ cardiac imaging, the current research is still far from‏ supporting clinical decision-making. Radiomics-based‏ cardiac imaging studies have been proved to show an‏ overall insufficient methodological quality (Lambin et‏ al. 2017; Ponsiglione et al. 2022). A more standardised‏ methodology in the radiomics workflow is needed to‏ cross the translational line between an exploratory‏ investigation method and a standardised added value to‏ precision medicine workflows.‏

Conflict of Interest

None.