Over 6 million individuals in the U.S. are diagnosed with heart failure and over 55 million individuals are diagnosed worldwide (1). Ischemic and non-ischemic cardiomyopathies are characterized by abnormalities in cardiac structure and function and frequently underlie heart failure. Presently, cardiovascular magnetic resonance imaging (CMR) is the gold standard non-invasive imaging method to evaluate for these abnormalities (2). CMR cine sequences can be used for qualitative or quantitative measures of ventricular structure and function including left ventricular (LV) mass, volume, or myocardial scarring; moreover, administration of gadolinium-based contrast agents in CMR sequences has been used to improve myocardial tissue characterization via gadolinium enhanced imaging. More recently, the use of parametric mapping, which generates a pixel-wise map of magnetic resonance properties of tissue (3), has enabled deeper characterization of the myocardium. Structural and tissue-level characterization of the myocardium and its pathology is particularly valuable in the evaluation of diffuse myocardial disease that is often characteristic of non-ischemic cardiomyopathy (NICM).

Radiomic analysis, which leverages advanced analysis of pixel- and voxel-level patterns of an image to extract a vast array of features, is a burgeoning field that enables more detailed and objective characterization of tissue structure and properties from medical images. Pioneered in the field of oncology where it has led to significant advances in tumor characterization (4), radiomics has numerous applications in cardiac imaging. Applied to CMR, radiomics can derive features of cardiac shape, myocardial texture, intensity, and higher-order statistics of pixel variations beyond traditional qualitative assessments or limited quantitative measures of ventricular volume and mass. It offers a particular promise in the diagnosis and evaluation of NICM, where tissue heterogeneity, fibrosis, and other diffuse abnormalities often elude conventional imaging assessment.

Advances in machine learning (ML) have enabled advanced computational image analysis to extend the capabilities of CMR in the diagnosis and characterization of NICM. Applying ML algorithms in the analysis of radiomic data offers a potential avenue to enhance diagnostic accuracy, prognostication, and treatment stratification for numerous forms and etiologies of NICM. In this review, we will provide an overview of the routinely acquired CMR sequences available for radiomic analysis, review its applications in the evaluation of NICM, and suggest areas for further opportunity in this field.

Non-invasive characterization of myocardial structure and function is an important step for determining the underlying etiology of a cardiomyopathy. Moreover, repeated evaluation is key to monitoring response to treatment and risk stratification for future complications like ventricular arrhythmia. CMR for cardiomyopathy includes T1-weighted, T2-weighted, and gadolinium-contrast enhanced cine imaging sequences as well as late gadolinium enhancement (LGE) imaging, well-regarded as the gold standards for the assessment fibrosis, edema, and infiltrative disease (5–7). However, traditionally, these modalities rely on visual assessment of signal intensity, thus limiting the ability to assess myocardial pathology that is uniform throughout the myocardium, such as diffuse myocardial inflammation, fibrosis, hypertrophy, or infiltration (8). In recent years, rapid parametric mapping techniques such as T1, T2, and T2* mapping, which yield quantitative data that captures various tissue-specific properties of the myocardium, have broadened the scope of tissue characterization and the ability for histologic inference (9, 10). Compared to T2 and T2*, T1 mapping has been the most widely used and broadly applicable for radiomic analysis, as well as the focus of many studies utilizing ML algorithms in radiomic data analysis.

T1 mapping generates a pixel-wise map of T1 relaxation times, an inherent property of the tissue reflecting its composition. Native T1 mapping, performed in the absence of GBCA, reflects the condition of both myocytes and the interstitial spaces, allowing for detection of myocardial abnormalities such as edema, protein deposition, and iron overload. Post-contrast T1 mapping, performed with GBCA, allows for an estimation of the extracellular volume (ECV) which aids in quantification interstitial fibrosis and infiltration. Thus, T1 mapping offers advantages in characterizing myocardial infarction, myocarditis, and systemic diseases such as amyloidosis and Fabry disease.

Using ML techniques with radiomic data from CMR has enhanced myocardial tissue characterization in NICM. By extracting quantitative features from standard imaging sequences, these techniques can improve diagnostic accuracy, aid in differential diagnosis, and potentially reduce the need for contrast agents. In this section, we review key studies that have leveraged radiomics and ML across various NICM subtypes. A summary of the referenced studies is provided in Table 2.

Further the additional integration of genotyping with radiomics has the potential to open new avenues in genetic prediction. One analysis demonstrated that radiomic analysis of native T1 mapping images could differentiate between MYH7 and MYBPC3 mutations HCM with high accuracy (AUC 0.88–0.90), suggesting CMR imaging capabilities beyond myocardial structure, function, and characterization and extending to genetic implications (15). The addition of radiomic texture analysis to predictive models of cardiovascular disease has also been shown to improve accuracy. Pujadas et al. compared models utilizing vascular risk factors, standard CMR features, and extracted radiomics features in various combinations to predict incident cardiovascular events, including atrial fibrillation, heart failure, and myocardial infarction (37); their analysis showed the addition of radiomic features to the model provided incremental predictive value over vascular risk factors and CMR indices. As radiomics research continues to identify imaging biomarkers associated with disease progression and outcomes, we foresee the evolution of more personalized treatment strategies. We also foresee that radiomic features may also assist in predicting both therapeutic and adverse reactions to medication, which may help guide clinicians towards therapeutic interventions optimized for individual patient risk profiles.

The field of radiomics in CMR has demonstrated remarkable potential for improving the evaluation and characterization of cardiac pathologies, particularly NICM. However, several challenges hinder widespread utilization, namely its interpretability, reproducibility, and generalizability.

First, interpretation of radiomic outputs and relating it to tissue pathology is nuanced. On a conceptual level, radiomic analysis characterizes the distribution and pattern of signal intensity within a segmented ROI and assumes that these patterns reflect some intrinsic nature of the tissue. With the improvement of this feature extraction and ML algorithms for this analysis, the goal is to generate radiomic signatures via unique patterns of signal intensity that can differentiate cardiac disease states and offer prognostic value. Inherent to this is the conversion of the qualitative image analysis to a mathematic quantification of imaging values that are harder to intuitively link to known tissue-level processes. This creates a kind of “black-box” effect as a model can make predictions based on features the clinician may not perceive and may create new challenges in interpretation.

Next, there are multiple ongoing challenges to reproducibility. While many of the studies reviewed herein are promising, variability in image acquisition protocols, scanner hardware, and reconstruction algorithms can introduce substantial heterogeneity in image quality and subsequently affect the reproducibility of extracted radiomic features and the generalizability of ML models. Establishing standardized imaging protocols, feature extraction methods, and analytical pipelines is essential before radiomics can reliably be brought to the clinical setting. The Image Biomarker Standardization Initiative has been a notable effort to establish guidelines for the standardization of radiomic feature definitions and extraction methods, promoting reproducibility and comparability across different research groups (38).

Lastly, most studies use relatively small, single-center datasets, limiting the generalizability of findings. Developing radiomics models for use in practice requires access to large, diverse datasets. However, data scarcity and patient privacy considerations necessarily constrain the availability and sharing of CMR data. Collaborative approaches are essential to overcome these challenges and enhance the generalizability of radiomics models. Large-scale databases like the UK Biobank, which includes imaging and clinical data from over 100,000 participants, have created extraordinary opportunities for cardiovascular radiomics research (39). Meanwhile, federated learning has emerged as an alternative solution to address data scarcity and privacy concerns in multi-center studies. In principle, federated learning allows ML models to be trained across multiple decentralized institutions holding local datasets, without the need to transfer or share raw patient data. While this solution has shown potential in cardiovascular radiomics (40), federated learning continues to draw concerns, particularly regarding coordination, expense, and handling of data heterogeneity (41, 42). Ultimately, multicenter collaborations can be expected to play a pivotal role in strengthening standardization efforts by fostering consensus on data formats, feature definitions, and validation methodologies.

The majority of radiomics studies in cardiology, including many reviewed here, have been retrospective and limited to small, single-center cohorts. Going forward, large-scale, prospective studies and randomized controlled trials should evaluate the impact of radiomics on clinical outcomes, cost-effectiveness, and patient quality of life. Such research will be essential in the eventual integration of radiomics into modern clinical practice.

However, integration of ML models into clinical practice raises several policy level, ethical, and regulatory considerations, which have increasingly fallen on the shoulders of regulatory bodies such as the Food and Drug Administration (FDA). The FDA has sought to reimagine a total product lifecycle approach for regulating AI and ML-based software as medical devices (43). This regulatory framework has sought to addressing health disparities by emphasizing good ML practices during premarket review and ongoing performance monitoring to mitigate group harms, however critics have raised concerns that these practices do not sufficiently address the potential for AI and ML models to perpetuate or exacerbate existing health disparities (44). Meanwhile, concerns regarding bias and transparency have been voiced among practitioners, particularly in radiology, where reliance on AI algorithms raises challenges such as failures in edge cases and the opacity of decision-making processes (45), highlighting potential gaps in current regulatory frameworks. Recently, the FDA has voiced the need for a harmonized global regulatory approach and advocated for transparent, risk-based regulatory strategies tailored to the complexity of AI systems, underscoring an ongoing effort to balance innovation and safety (46).

To address challenges, several solutions have been proposed. Standardization of imaging protocols and harmonization of data processing pipelines are crucial steps in reducing variability and improving reproducibility. Techniques such as ComBat or ML-based normalization can mitigate biases introduced by differences in scanners and protocols. Collaborative efforts to create large, open-access CMR datasets with standardized annotations are essential for enabling robust training and validation of radiomics models. Federated learning offers an innovative approach to facilitate multi-center collaborations while preserving data privacy. Advances in explainable AI can improve the interpretability of radiomic features, allowing clinicians to better understand their relevance to cardiac pathophysiology. Combining radiomics with other data sources, such as proteomics and genomics, could further enhance its clinical utility. Moreover, tissue-based pathologic validation would allow for a deeper understanding of the relationship between radiomics features and histopathological disease correlates. Streamlining regulatory pathways for AI-based tools and fostering collaborations between academia, industry, and regulatory bodies would accelerate the translation of radiomics into clinical practice. Moreover, education and training programs for clinicians and radiologists on radiomic principles and applications can increase acceptance and integration of these tools into workflows. Development of user-friendly software platforms for automated feature extraction would further ease their adoption. Looking ahead, the integration of radiomics with other advanced imaging techniques, such as strain imaging and T1/T2 mapping, as well as multi-omics data, presents exciting opportunities to provide a more comprehensive understanding of cardiac pathophysiology. Radiomics has the potential to identify subclinical myocardial changes before overt disease manifests, paving the way for earlier interventions and personalized therapies. Future efforts should also focus on leveraging radiomics for prognostication and therapy monitoring, particularly in conditions such as HCM and CA. Advances in computational efficiency and cloud-based processing could enable real-time radiomic analysis, providing immediate insights during clinical encounters. Ultimately, fostering interdisciplinary collaboration between clinicians, data scientists, and imaging physicists will be critical to addressing the multifaceted challenges of radiomics and ensuring its clinical relevance. By addressing current limitations and embracing innovative solutions, radiomics in CMR can become a transformative tool for advancing precision cardiology.