Artificial intelligence in cardiovascular magnetic resonance imaging

Abstract

Artificial intelligence is rapidly evolving and its possibilities are endless. Its primary applications in cardiac magnetic resonance imaging have focused on: image acquisition (in terms of acceleration and quality improvement); segmentation (in terms of saving time and reproducibility); tissue characterisation (including radiomic techniques and the non-contrast assessment of myocardial fibrosis); automatic diagnosis; and prognostic stratification. The aim of this article is to attempt to provide an overview of the current situation as preparation for the significant changes currently underway or imminent in the very near future.

Resumen

La Inteligencia artificial está en pleno desarrollo y sus posibilidades son incalculables. Sus principales aplicaciones en la resonancia magnética cardiaca se han focalizado en: la adquisición de las imágenes (en términos de aceleración y mejoría de la calidad); la segmentación (en términos de ahorro de tiempo y reproducibilidad); la caracterización tisular (incluyendo técnicas de Radiomica y valoración de la fibrosis miocárdica sin contraste); en el diagnóstico automático y en la estratificación pronóstica. El objetivo de este artículo es intentar dar una visión de la situación actual, como preparación para los grandes cambios en curso o que vendrán a muy corto plazo.

Introduction

In recent years, cardiac imaging techniques have undergone numerous changes and improvements. Despite these advancements, the heart remains a challenging organ to study due to its intrinsic characteristics, including its complex three-dimensional (3D) movement, significant pathophysiological variability and continuous respiratory motion. Furthermore, there are multiple imaging protocols to choose from depending on the clinical context. Each imaging study must assess multiple structures while taking into account both macroscopic and microscopic elements, the complex ventricular/atrial geometry, the great vessels and the wide variability present among patients.

These factors represent just one aspect of the many complexities inherent in cardiac imaging. Cardiac magnetic resonance imaging (MRI) is one of the fastest-growing non-invasive imaging techniques, with an increasing number of indications due to its broad application in diagnosis, prognosis, and prevention.^1,2 Considering that cardiovascular disease is currently the leading cause of morbidity and mortality worldwide,³ coupled with the growing trend toward defensive medicine and the lengthy duration of cardiac MRI studies, it is evident that our medical and diagnostic system is poised for significant transformation.

Artificial intelligence (AI) is emerging as one of the most significant innovations and potential solutions in this context, particularly due to its impact on multiple aspects of imaging. This article focuses on AI applications in cardiac MRI, specifically in areas where it is most relevant: image acquisition, segmentation, tissue characterisation, and its contribution to diagnosis and prognosis.

Several definitions of AI are currently in use. It is generally described as the ability of a computer or machine to autonomously ‘learn’ from experience and replicate human intelligence processes to solve a broad range of tasks.^4,5 In the medical field, it is widely held that two primary factors define the skilled professional: knowledge and experience. The primary constraint on the human mind’s ability to absorb large amounts of data is time. In contrast, software can leverage modern algorithms to acquire far greater ‘experience’ than the human mind in far less time.⁶

AI encompasses several concepts and techniques including the key subcategories of machine learning (ML) and deep learning (DL).⁷

ML was developed to create algorithms capable of making decisions and learning from the outcomes. In practice, these systems can learn without explicit prior programming. ML algorithms can be subdivided into supervised, unsupervised and reinforced learning.^5,7 Supervised learning techniques rely on correctly labelled datasets where the correct output for each data point is provided. In contrast, unsupervised learning techniques autonomously identify hidden patterns or structures within data without any prior knowledge of the desired or potential output. This method is potentially capable of discovering new relationships and/or previously unknown clusters. Reinforced learning algorithms learn to perform specific tasks through a reward/penalty system (positive and negative reinforcement). In cardiovascular imaging, common ML techniques include logistic regression, support vector machine, random forest, cluster analysis and the use of both artificial and convolutional neural networks.^5,7

Representation learning is a subset of ML in which algorithms autonomously extract multiple features to classify available data. DL, in turn, is a subcategory of representation learning.⁸ DL is characterised by a multi-layered processing structure that uses artificial neural networks inspired by biological neural networks. This architecture enables multi-level learning with the capacity to learn complex functions without relying on specific algorithms.^5,7 The greater the data volume, the higher the number of layers required. These layers can be categorised into three main compartments: input layer, hidden layer and output layer. Current models allow human control over input and output layers, but not hidden layers, which are thus referred to as ‘black boxes’⁷ (Fig. 1).

Figure 1.

Example of a Convolutional Neural Network structure. Source: Fotaki et al.¹⁰

Another frequently mentioned AI-related term in cardiac MRI is convolutional neural network (CNN). This type of DL architecture is designed to analyse image or video data, extracting multiple features simultaneously from the available inputs.^9–11 Most DL techniques are based on CNNs, with U-Net being the most widely used architecture for detection, localisation, and segmentation.^4,12,13

Principal applications in cardiac magnetic resonance imaging

Image acquisition

Image acquisition times have long been one of the main limitations of cardiac MRI with scanning times typically ranging from 30−60 min, depending on the clinical context. Additionally, given their importance in clinical practice, cardiac MRI images must provide high spatial, temporal, and contrast resolution while thoroughly assessing all cardiac structures. These requirements necessitate relatively long acquisition times due to the need for cardiac and respiratory synchronisation techniques. As a result, there has been a long-standing effort to enhance acquisition speed by introducing innovative techniques and sequences, such as pulse sequences, motion correction approaches, parallel reconstruction, or compressed sensing. However, the main limitation of compressed sensing at high acceleration parameters is the increased complexity of algorithms, prolonged reconstruction times, and potential image degradation.¹¹

AI offers significant advantages in reducing scanning time, benefiting both patients (shorter studies, more manageable breath-holds, etc.) and the healthcare system. ML-based algorithms have been proposed for the non-linear optimisation of image reconstruction.¹¹ The primary goal is to simplify acquisition, reduce scanning times and enhance efficiency, while in some cases improving image quality.^7,14 For instance, Bustin et al.¹¹ achieved a significant acceleration in image acquisition using an under-sampling strategy followed by an ‘experience-based’ estimation of the remaining domain based on the acquired data. This outcome was made possible with DL techniques, which enabled the offline transformation (based on training data) of low-resolution images into high-resolution images. In this context, AI—and particularly DL techniques— can shift the complexity of reconstructions previously performed in-line to an offline training process.¹¹

Among other examples, Küstner et al.¹⁵ used DL-based methods to develop a 3D cine acquisition (CINENet) in a single 10–15-second breath-hold with an acceleration factor of approximately 10–15 × . In another study using DL techniques, Küstner et al.¹⁶ obtained super-resolution coronary MR angiography images (isotropic voxel of 1.2 mm) from a low-resolution acquisition (1.2 × 4.8 × 4.8 mm). This was achieved using a 50-second free-breathing acquisition, resulting in a 16x increase in spatial resolution. Similar results, with a 9× acceleration, were obtained by Fuin et al.¹⁷

Using a U-Net algorithm, Hauptman et al.¹⁸ successfully eliminated aliasing artifacts and calculated cardiac volumes in cine images 13 times faster than conventional acquisitions.

Several commercial manufacturers have also employed AI to enhance image quality, leading to relative increases in acquisition speed. Notable examples include the AIR Recon DL (GE Healthcare), a DL-CNN-based algorithm that produces high-resolution, low-noise images while improving sharpness and reducing truncation or Gibbs artifacts.¹⁹ Siemens Healthineers (Erlangen, Germany) has developed a DL-based technique called Deep Resolve which is further characterised by two sub-algorithms: Deep Resolve Gain and Deep Resolve Sharp. Together, they can decrease acquisition time while improving image quality.^18,20SmartSpeed is an AI method developed by Philips which applies a DL-CNN algorithm called Adaptive-CS-Net to compressed sensing techniques (artificial intelligence compressed sensing [AICS]) which can be used in both 2D and 3D Cartesian acquisitions, achieving 30–50% acceleration without compromising quality.²¹ (Figs. 2 and 3).

Figure 2.

Comparison of short-axis cine SSFP images in diastole (top row) and systole (bottom row). Left column: SENSE (voxel size 8 × 1.8 × 1.8 mm), acquisition time per slice: 16 s. Centre column: Compressed SENSE (voxel size 8 × 1.8 × 1.8 mm), acquisition time per slice: 6 s. Right column: Compressed SENSE with AI (AICS - SmartSpeed), (voxel size 8 × 1.8 × 1.8 mm), acquisition time per slice: 4 s. Acceleration parameters optimised to achieve similar image quality with same signal-to-noise ratio. Despite accelerated acquisition of images with AI (right column), right ventricular endocardial trabeculae are better defined. Study performed using Philips Ingenia 1.5 T scanner (Philips Healthcare, Best, Netherlands) at Hospital Universitario Nuestra Señora del Rosario, Madrid, Spain.

AICS: artificial intelligence compressed sensing; AI: artificial intelligence; SSFP: steady-state free precession.

Figure 3.

3D white-blood acquisition of thoracic aorta with ECG and respiratory gating. Left: compressed sensing. Estimated duration: two minutes. Right: compressed sensing combined with AI (AICS - SmartSpeed), same voxel size. Estimated duration: 50 seconds. The right image demonstrates superior quality with optimal visualisation of the coronary origin (yellow arrows: right coronary artery) due to improved patient tolerance to shorter acquisition time and reduced impact of respiratory irregularities. Study performed using Philips Ingenia 1.5 T scanner (Philips Healthcare, Best, Netherlands) at Hospital Universitario Nuestra Señora del Rosario, Madrid, Spain.

AICS: artificial intelligence compressed sensing; ECG: electrocardiography; AI: artificial intelligence.

Another way of optimising scanning time in cardiac MRI studies is to reduce the time needed for sequence planning. AI can also assist operators by automatically recognising reference points, thereby reducing waiting and planning times while also improving intra- and inter-operator reproducibility.^22,23

Segmentation

Manual or semi-automatic segmentation remains the standard practice for calculating morphological and functional parameters for cardiac chambers in most centres. This process is labour-intensive and sensitive to intra/inter-observer variability, a significant limitation when considering that MRI is the gold standard for assessing cardiac morphology and function.²⁴ In recent years, increasingly effective and high-quality AI-based automatic segmentation techniques have been introduced, allowing for faster and more reproducible segmentation than the manual approach, overcoming classic issues associated with the most critical regions (the basal areas of both ventricles). Notably, Karimi-Bidhendi et al. developed a DL-CNN-based algorithm, specifically a fully convolutional network, to achieve fully automatic segmentation in paediatric patients with complex congenital heart disease.²⁵ Davies et al. demonstrated in 2022 that AI can surpass human capabilities, performing left ventricular segmentation more quickly (20 s vs 13 min per patient) and with greater accuracy.²⁶ Another group has reported that cardiac measurements obtained through right heart catheterisation correlate better with those obtained via DL than those obtained through manual segmentation.²⁷

Other methods, many of them based on DL and a few on ML, have also been studied. One interesting method based on the former is described in the work of Alandejani et al.²⁸ This method was trained to perform automatic segmentation of the right atrium (size and function), demonstrating excellent reproducibility with an intraclass correlation coefficient (ICC) of 0.91–0.95, compared to manual measurements of 0.82–0.88 and 0.88–0.91 by two different operators. It also demonstrates a moderate association with invasive haemodynamic studies such as the correlation of mean right atrial pressure with the automatic measurement of the maximum area (r = 0.64 and minimum r = 0.66) compared to manual measurements (r = 0.57), all with p < 0.001. Furthermore, it can predict mortality in patients with pulmonary hypertension, with the minimum area measurement showing a hazard ratio of 1.02 and a 95% confidence interval of 1.01–1.03.²⁸ There are also aortic segmentation methods, whole-heart segmentation, and even automatic segmentation of areas of myocardial fibrosis/scars.^7,29 Among the latter, it is also worth mentioning the work of Xu et al., who described a DL method for segmenting infarct areas in non-contrast cine images.³⁰

Bratt et al. published an ML method for fully automatic segmentation of aortic flow quantification in phase contrast sequences, with minimal differences compared to manual segmentation and significant time savings (<0.01 min per case compared to the manual average of four minutes per case).³¹ Another study explored using CNN-based algorithms to fully automate the segmentation of myocardial strain cine sequences using Displacement Encoding with Stimulated Echoes (DENSE).³²

4 D flow sequences have been widely studied in recent years due to their significant clinical implications in multiple contexts, ranging from congenital heart disease and shunts to aortic dilations. However, segmentation is often labour-intensive, making it difficult to integrate into clinical practice at most centres. Berhane et al. and other groups have investigated automatic segmentation techniques using DL-CNN algorithms.³³ These studies could help introduce this tool into daily practice and expand its use.

Tissue characterisation

The administration of gadolinium and the assessment of late gadolinium enhancement (LGE) remain fundamental pillars of cardiac MRI, playing a crucial role in several processes including diagnosis and prognostic stratification.^34,35 AI has been utilised in this domain for the automatic quantification of myocardial scar in ischaemic heart disease, as demonstrated by Zabihollahy et al.³⁶ and Popescu et al.³⁷ as well as in non-ischaemic heart disease³⁸ and even within the atria.³⁹

However, despite its significant benefits, LGE assessment is not without risks, and can be problematic in patients with contraindications (clearance < 30 mL/min, allergies, etc.). AI advancements in this context have enabled scar detection in various cardiopathies without the need for intravenous contrast.^30,40,41 In particular, Zhang et al.⁴⁰ developed a CNN-based virtual native enhancement (VNE) algorithm capable of producing images without gadolinium administration that are similar to those generated with conventional LGE from T1 maps and cine sequences. Studies have explored how this method could replace LGE imaging for scar assessment in patients with hypertrophic cardiomyopathy (n = 1348), demonstrating a high correlation with conventional LGE sequences both in hyperintense lesions (r = 0.77–0.79; ICC = 0.77–0.87; p < 0.001) and in intermediate-intensity lesions (r = 0.70–0.76; ICC = 0.82–0.85; p < 0.001), with the VNE images having a better objective image quality⁴⁰ (Fig. 4). They subsequently used this method in patients with ischaemic heart disease, obtaining similarly high correlation with LGE imaging in addition to better image quality.⁴¹

Figure 4.

CNN-based virtual native enhancement (VNE) from T1 maps and cine sequences without intravenous (iv) gadolinium; compared to conventional late gadolinium enhancement (LGE). Application in patients with hypertrophic cardiomyopathy. Source: Zhang et al.⁴⁰ Open access article through Commons Attribution License (CC BY) (https://creativecommons.org/licences/by/4.0/.

CNN: Convolutional Neural Network; iv: intravenous.

In the broad field of tissue characterisation, it is essential to consider developments in radiomics—an ML technique designed to extract high-dimensional data. This applies to the analysis of medical images to obtain quantitative information that would be impossible to acquire through simple visual assessment.⁷ One of the most recent techniques is texture analysis (TA), which uses various ML algorithms to quantify spatial heterogeneity and the relationship between different pixels. This analysis is based on the idea that the distribution of grey levels across voxels provides much richer information than merely measuring its signal. The implementation of this technique thus enables a shift from qualitative to quantitative assessment.¹⁰

Baessler et al. studied the application of TA to non-contrast cine sequences, achieving the detection of ischaemic scars in both extensive and smaller infarctions with high diagnostic accuracy (area under the curve [AUC] of 0.93 and 0.92, respectively).⁴² Similar results were also published by Avard et al.⁴³ Meanwhile, Neisius et al.⁴⁴ were able to differentiate patients with hypertensive heart disease from those with hypertrophic cardiomyopathy, conditions that are often difficult to differentiate even with MRI. They achieved this result by analysing native T1 mapping sequences using TA which attained a diagnostic accuracy of 86.2% compared to standard native T1 mapping assessment which attained an accuracy of 64%. Mancio et al.⁴⁵ also employed radiomics in the context of hypertrophic cardiomyopathy, identifying the presence of myocardial scar in cine sequences with good results (sensitivity of 91%, specificity of 62%, and a negative predictive value [NPV] of 89%). Other studies have noted that TA can differentiate cardiac amyloidosis from hypertrophic cardiomyopathy in non-contrast cine sequences.⁴⁶

These advancements could lead to a significant change to the gadolinium/fibrosis paradigm, impacting how cardiac MRI evolves, with its advantages including the optimisation of scanning times, reduction of gadolinium administration, and the possibility to carry out studies on patients with contrast contraindications.

Applications in diagnosis

Studies on AI in cardiac MRI have focused primarily on acquisition, segmentation and the estimation of morphological and functional cardiac indicators. However, the potential applications of AI extend beyond these areas and could have a significant clinical impact on decision-making in the future as AI facilitates early diagnosis and predicts treatment responses. AI-based methods have been developed to diagnose conditions through cardiac MR image analysis.^7,47 Initially, ML methods were the most widely used due to their favourable performance; however, they have proved to be inefficient given their high computational complexity. In this regard, DL algorithms have managed to overcome the limitations of ML.⁴⁷ To date, numerous studies have been published regarding this application of AI, largely facilitated by the multiple datasets available to researchers.⁴⁷

In 2020, Martin-Isla et al. conducted a detailed review of ML applications in cardiac MRI diagnostics, analysing their advantages and limitations.⁴⁸ More recently, in 2023, Jafari et al. carried out an extensive review focusing exclusively on DL methods for automated diagnosis, analysing over 230 articles published between 2016 and 2022.⁴⁷ Within this extensive body of literature, several groups have achieved high diagnostic accuracy in pathological classification, distinguishing between normal patients and those with hypertrophic cardiomyopathy, dilated cardiomyopathy, ischaemic heart disease or right ventricular abnormalities. In one such study, Ammar et al. attained an accuracy of 0.92 when analysing cine images using CNN-U-Net methods.⁴⁹ Similarly, other researchers including Snaauw et al.⁵⁰ have reported good results in classifying the same pathologies.

Applications in prognosis

Among the various potential applications of AI, the prognostic stratification of patients is particularly relevant. For instance, ML algorithms have been applied to a large number of variables (n = 735), including cardiac MRI data for 6814 participants in the Multi-Ethnic Study of Atherosclerosis (MESA), where an improvement was observed in the accurate prediction of cardiovascular events in initially asymptomatic patients.⁵¹

It is well established that the quantification of myocardial scar in terms of core and peri-infarct zone on cardiac MRI is a predictive factor for sudden cardiac death.⁵² In recent years, there has been an ever-increasing recognition of the prognostic significance of scar characteristics (morphology, microstructure, heterogeneity), beyond a sole focus on scar extent.⁵³ However, analysing these characteristics is complex and time-consuming, making it impractical for routine clinical practice. Zaidi et al.⁵³ used ML techniques to characterise myocardial scar microstructure in LGE sequences. In this study, they achieved improved risk stratification (predicting major arrhythmic events) compared with conventional predictors (Fig. 5). In another multicentre study analysing more than 700 patients, fully automated ML techniques for myocardial scar mass quantification resulted in an increased capacity to predict major cardiac events.⁵⁴

Figure 5.

Method for extracting microstructure features in scar areas on late gadolinium enhancement (LGE) images using machine learning (ML) algorithm for arrhythmic risk stratification. Source: Zaidi et al.⁵³ Open access article through Commons Attribution License (CC BY) (https://creativecommons.org/licences/by/4.0/).

ML: machine learning.

A further example of the emerging role of AI in prognostics was published by Seraphim et al. in 2021.⁵⁵ Their study developed an ML algorithm for the automatic calculation of pulmonary transit time (and its derived parameter, pulmonary blood volume index) in perfusion studies through cardiac MRI. Their results confirmed the potential value of these measurements as independent predictors of adverse cardiac events, supporting their integration into clinical practice.⁵⁵ In another study, mortality in patients with pulmonary hypertension was predicted using automatic segmentation of the right ventricle.²⁷ Finally, among its many other applications, AI has been employed in cardiology (including cardiac MRI and other parameters) to predict post-cardiac resynchronisation therapy survival.⁵⁶

Limitations, future perspectives, and conclusions

The future of AI in the field of cardiac MRI appears virtually limitless. However, its use continues to raise concerns in the medical community, particularly among radiologists and cardiologists. One of the primary issues is the reliability of DL algorithm outputs, given the ‘black-box’ hidden-processing layers structure of these algorithms. This design makes it difficult to understand the mechanisms by which these algorithms generate their results, thereby complicating the identification of discrete errors and raising ethical concerns regarding the entrusting of patient health to an ‘unknown’ process. Additionally, AI relies on experience gained from training data, meaning its performance is limited for rare diseases on which it has not been trained. Another limitation to advancing AI in cardiac MRI is the significant variability in imaging techniques, protocols, sequences and manufacturers as well as population and racial diversity worldwide, complicating large-scale implementation. Other concerns relate to patient data privacy and security. Finally, it is important to remember that every imaging diagnosis must be considered within the context of its clinical history, laboratory findings, and patient background. Therefore, image analyses in the absence of a clinical context could likely lead to inaccurate or ineffective conclusions. Fenech et al.⁵⁷ have published an interesting article on the ethical concerns surrounding AI in imaging studies.

In conclusion, AI is rapidly evolving and experiencing exponential growth, making its adoption inevitable and, above all, necessary in the current era. It is crucial to understand its advantages and applications, but in the medical field, it is perhaps even more important to recognise its limitations.

CRediT authorship contribution statement

• 1

  Research coordinators: A. Castellaccio and E. Vañó Galván

• 2

  Study concept: A. Castellaccio

• 3

  Study design: A. Castellaccio

• 4

  Data collection: A. Castellaccio, N. Almeida Arostegui and E. Vañó Galván

• 5

  Data analysis and interpretation: -

• 6

  Data processing: -

• 7

  Literature search: A. Castellaccio, N. Almeida Arostegui, M. Palomo Jiménez, D. Quiñones Tapia and M. Bret Zurita

• 8

  Drafting of article: A. Castellaccio, E. Vañó Galván and N. Almeida Arostegui

• 9

  Critical review of the manuscript with intellectually relevant contributions: E. Vañó Galván, N. Almeida Arostegui, M. Palomo Jiménez, D. Quiñones Tapia and M. Bret Zurita

• 10

  Approval of the final version: A. Castellaccio and E. Vañó Galván