The Applications of Artificial Intelligence in Cardiovascular Magnetic Resonance—A Comprehensive Review

Cardiovascular disease remains an integral field on which new research in both the biomedical and technological fields is based, as it remains the leading cause of mortality and morbidity worldwide. However, despite the progress of cardiac imaging techniques, the heart remains a challenging organ to study. Artificial intelligence (AI) has emerged as one of the major innovations in the field of diagnostic imaging, with a dramatic impact on cardiovascular magnetic resonance imaging (CMR). AI will be increasingly present in the medical world, with strong potential for greater diagnostic efficiency and accuracy. Regarding the use of AI in image acquisition and reconstruction, the main role was to reduce the time of image acquisition and analysis, one of the biggest challenges concerning magnetic resonance; moreover, it has been seen to play a role in the automatic correction of artifacts. The use of these techniques in image segmentation has allowed automatic and accurate quantification of the volumes and masses of the left and right ventricles, with occasional need for manual correction. Furthermore, AI can be a useful tool to directly help the clinician in the diagnosis and derivation of prognostic information of cardiovascular diseases. This review addresses the applications and future prospects of AI in CMR imaging, from image acquisition and reconstruction to image segmentation, tissue characterization, diagnostic evaluation, and prognostication.


Introduction
Cardiovascular disease remains an integral field on which new research in both the biomedical and technological fields is based, as it remains the leading cause of mortality and morbidity worldwide [1]. Despite the progress of cardiac imaging techniques, the heart remains a challenging organ to study, secondary to respiratory motion, continuous cycles

Logistic Regression
Logistic regression is a supervised machine learning technique. It is based on the use of the logistic function (sigmoid), which converts real values to a value between 0 and 1.

Logistic Regression
Logistic regression is a supervised machine learning technique. It is based on the use of the logistic function (sigmoid), which converts real values to a value between 0 and 1. In the training phase, the algorithm receives a training dataset consisting of N examples. Each example consists of attributes X and a label Y indicating the correct classification. At the end of the training, the algorithm produces a model that can be used to classify any other example not included in the training set.

Support Vector Machine
Support vector machines (SVMs) are supervised ML models segregating the data into two or more classes, obtaining a linear, binary, nonprobabilistic classifier. An SVM model is a representation of the examples as points in space, mapped in such a way that the examples belonging to the two different categories are separated by as large a space as possible. New examples are then mapped in the same space, and the prediction of the category to which they belong is made based on the side in which they fall. In addition to linear classification, it is possible to make use of SVMs to effectively carry out nonlinear classification using the kernel method by implicitly mapping their inputs into a multidimensional characteristics space.

Random Forest
The random forest (RF) method is an ensemble learning method for classification, a regression that operates by building a multitude of decision trees during training. For classification activities, the random forest output is the class selected by most trees. For regression activities, the mean or mean forecast of the individual trees is returned.

Cluster Analysis
Cluster analysis finds subgroups within the input data in an unsupervised manner. Clustering algorithms group elements based on their mutual distance, and therefore, whether or not they belong to a set depends on how far the element under consideration is from the set itself.

Artificial Neural Network
Artificial neural network (ANN) is a computational model composed of artificial "neurons", vaguely inspired by the simplification of a biological neural network, determined by the propagation of input data through a nonlinear transformation network.

Convolutional Neural Network
Convolutional neural networks (CNN) are the backbone of DL applications. A convolutional neural network is a type of feed-forward artificial neural network in which the connectivity pattern between neurons is inspired by the organization of the animal visual cortex, the individual neurons in which are arranged in such a way as to respond to the overlapping regions that tessellate the visual field.

Deep Learning
Deep learning (DL) evolved from ML ( Figure 2). It works in a layered architecture and uses the artificial neural network, a concept inspired by the biological neural network. It simply takes the data connections among all the artificial neurons and adjusts them based on the data model. More neurons are needed if the data size is large. It automatically presents learning at multiple levels of abstraction, thus allowing a system to learn the mapping of complex functions without depending on any specific algorithm. In DL, the accuracy of the output depends on the amount of data. It consists of three layers:

1.
Input layer: the input layer is used to take input data from sources and then pass it to the hidden layers of the neural network. It does not perform any calculations.

2.
Hidden level: this level consists of many hidden levels. All the calculation is performed at this level. After all the calculations are complete, it proceeds to the output level.

3.
Output level: this level is used to provide the output to the outside world.
It simply takes the data connections among all the artificial neurons and adjusts them based on the data model. More neurons are needed if the data size is large. It automatically presents learning at multiple levels of abstraction, thus allowing a system to learn the mapping of complex functions without depending on any specific algorithm. In DL, the accuracy of the output depends on the amount of data. It consists of three layers: 1. Input layer: the input layer is used to take input data from sources and then pass it to the hidden layers of the neural network. It does not perform any calculations. 2. Hidden level: this level consists of many hidden levels. All the calculation is performed at this level. After all the calculations are complete, it proceeds to the output level. 3. Output level: this level is used to provide the output to the outside world.

Current Applications of Artificial Intelligence
Machine learning will have an impact on all aspects of CMR, from patient programming to image analysis and prognosis. Optimal CMR imaging requires correct patient positioning and precise image planning [26]. The acquisition and reconstruction of images is a task typically entrusted to human experience that can now be automated with AI and reduce time. Machine learning methods have been used to optimize frequency regulation for 3-Tesla CMR and for automatic correction of artifacts. Recently, DL techniques have been shown to offer superior performance in terms of reconstruction quality as well as offering high efficiency [27]. Although CMR imaging offers many benefits for assessing cardiac structure and function, CMR image acquisition is time

Current Applications of Artificial Intelligence
Machine learning will have an impact on all aspects of CMR, from patient programming to image analysis and prognosis. Optimal CMR imaging requires correct patient positioning and precise image planning [26]. The acquisition and reconstruction of images is a task typically entrusted to human experience that can now be automated with AI and reduce time. Machine learning methods have been used to optimize frequency regulation for 3-Tesla CMR and for automatic correction of artifacts. Recently, DL techniques have been shown to offer superior performance in terms of reconstruction quality as well as offering high efficiency [27]. Although CMR imaging offers many benefits for assessing cardiac structure and function, CMR image acquisition is time consuming because of complex cardiac and respiratory movements. Thus, reducing scan time has always been one of CMR imaging's biggest challenges. Over the past decade, methods such as parallel imaging and compressed sensing (CS) have been increasingly used to accelerate CMR image acquisition (Figures 3 and 4    . Short axis acquired using parallel imaging and compressed sensing method shows a functional cardiac plane acquired using parallel imaging with a 1.5T MR syste string shows an SA acquired using the CS method with a 3T MR system. SA-sho compressed sensing.

Image Acquisition
Several AI solutions have been proposed by different vendors in terms of image acquisition and reconstruction (Table 1) [31,32]. The main goals of all these approaches are to simplify the image acquisition, facilitate CMR acquisition, and in some cases improve image quality, reduce the time of acquisition, and improve overall efficiency [27]. Forman et al. studied a scan protocol for coronary magnetic resonance angiography (CMRA) based on multiple breath holds featuring 1D motion compensation and compared the resulting image quality with that from a navigator-gated free-breathing acquisition. In other work, the investigators used iterative reconstruction with self-navigated free-breathing CMRA for retrospective reduction of respiratory motion artifacts ( Figure 5) [33,34]. Nakamura et al. demonstrated that noncontrast CMRA using CS could greatly shorten acquisition time compared with that of conventional navigator-gated coronary MRA while maintaining acceptable visualization at 3T [35].

Image Acquisition
Several AI solutions have been proposed by different vendors in terms o acquisition and reconstruction (Table 1) [31,32]. The main goals of all these app are to simplify the image acquisition, facilitate CMR acquisition, and in som improve image quality, reduce the time of acquisition, and improve overall efficie Forman et al. studied a scan protocol for coronary magnetic resonance angio (CMRA) based on multiple breath holds featuring 1D motion compensati compared the resulting image quality with that from a navigator-gated free-b acquisition. In other work, the investigators used iterative reconstruction w navigated free-breathing CMRA for retrospective reduction of respiratory motion ( Figure 5) [33,34]. Nakamura et al. demonstrated that noncontrast CMRA using C greatly shorten acquisition time compared with that of conventional navigat coronary MRA while maintaining acceptable visualization at 3T [35].

Slice Position
Automatic slice positioning was one of the first applications of AI during image acquisition [36]. Frick et al. described a fully automated approach in which slice position was detected using a deformation and segmentation algorithm. Subsequently, several algorithms have been developed that were able to identify cardiac landmarks and automatically plan the cardiac planes [37][38][39]. Lu et al. proposed an algorithm that was able to identify landmarks and prescribe long and short axis views from a 3D acquisition [40]. The authors created a model that was subsequently adapted to the heart of the patient and allowed prescribing the slices [40]. A more recent approach was proposed by Blansit et al., who developed an algorithm based on DL training on hundreds of annotated landmarks on CMR [41]. These approaches, based on training of 2D datasets, represent the basis for several vendors and may allow the ability to prescribe planes with low variability compared with those prescribed via manual acquisition.

Image Quality
Improvement of image quality has been developed by several vendors by decreasing the noise of pictures ( Figure 6). The "Deep Resolve" algorithm is a deep learning reconstruction able to decrease the noise of images. It is divided into "sharp" and "gain" reconstructions. In "Deep Resolve Gain" acquisition, the algorithm, starting from raw data, identifies the parts of images with more noise and increases the noise reduction in these parts [47]. In "Deep Resolve Sharp" acquisition the images enter the image reconstruction in a "high-resolution mode" while they are acquired with low resolution. The algorithm was trained on high-resolution images, so in the presence of low resolution pictures, the algorithm decreases the noise, simulating an image with high resolution from the raw data in order ensure the quality. these parts [47]. In "Deep Resolve Sharp" acquisition the images enter the image reconstruction in a "high-resolution mode" while they are acquired with low resolution.
The algorithm was trained on high-resolution images, so in the presence of low resolution pictures, the algorithm decreases the noise, simulating an image with high resolution from the raw data in order ensure the quality. , and NR 50% (C), respectively. The increasing percentage in NR reconstruction yielded a progressive reduction in image noise in 2D-MSLGE starting from NR 0% (C) and moving through NR 25% (D) and NR 50% (E). A breath artifact characterizing the inferior and inferolateral midapical segments was reduced in the reconstruction in which the 100% artificial intelligence algorithm was applied. In fact, a reduction in quantum noise resulted in better contrast resolution. 2D-MSLGE-2D multisegment late gadolinium enhancement; NR-artificial intelligence reconstruction deep learning noise reduction.
AIR Recon DL reconstruction is another technique that is able to increase signal-tonoise ratio, reduce truncation artifacts, and increase spatial resolution compared with standard reconstruction [42]. Using the AIR DL reconstruction, images can be reconstructed using "low", "medium", or "high" filters. The algorithm was created using millions of images, gradient backpropagation, and the ADAM optimizer. Compared with the previous algorithm, AIR DL has been evaluated in CMR [32]. Van der Velde et al. analyzed the impact of AIR DL on late gadolinium enhancement (LGE) images, demonstrating that it was possible to obtain LGE images with decreased noise [43]. Similar results were observed by Muscogiuri et al., who showed that it was possible to obtain images from multisegment LGE with the same image quality as those from standard single-segment LGE [32].

Image Speed Acquisition
Cine images in CMR result from a compromise between temporal and spatial resolution; in particular, decreased time of acquisition may cause image quality impairment [48]. However, compressed sensing provides the possibility of decreasing the time of cine image acquisition [49]. Hauptman et al. trained a U-Net algorithm that was able to remove aliasing and provide similar volumes in cine images thirteen times faster than standard cine images [44]. Sandino et al. developed another model, called "DL-ESPIRIT", that was able to combine the CNN reconstruction with standard reconstruction and provide similar results to standard CS images, accelerating the time of acquisition by , and NR 50% (C), respectively. The increasing percentage in NR reconstruction yielded a progressive reduction in image noise in 2D-MSLGE starting from NR 0% (C) and moving through NR 25% (D) and NR 50% (E). A breath artifact characterizing the inferior and inferolateral midapical segments was reduced in the reconstruction in which the 100% artificial intelligence algorithm was applied. In fact, a reduction in quantum noise resulted in better contrast resolution. 2D-MSLGE-2D multisegment late gadolinium enhancement; NR-artificial intelligence reconstruction deep learning noise reduction.
AIR Recon DL reconstruction is another technique that is able to increase signalto-noise ratio, reduce truncation artifacts, and increase spatial resolution compared with standard reconstruction [42]. Using the AIR DL reconstruction, images can be reconstructed using "low", "medium", or "high" filters. The algorithm was created using millions of images, gradient backpropagation, and the ADAM optimizer. Compared with the previous algorithm, AIR DL has been evaluated in CMR [32]. Van der Velde et al. analyzed the impact of AIR DL on late gadolinium enhancement (LGE) images, demonstrating that it was possible to obtain LGE images with decreased noise [43]. Similar results were observed by Muscogiuri et al., who showed that it was possible to obtain images from multisegment LGE with the same image quality as those from standard single-segment LGE [32].

Image Speed Acquisition
Cine images in CMR result from a compromise between temporal and spatial resolution; in particular, decreased time of acquisition may cause image quality impairment [48]. However, compressed sensing provides the possibility of decreasing the time of cine image acquisition [49]. Hauptman et al. trained a U-Net algorithm that was able to remove aliasing and provide similar volumes in cine images thirteen times faster than standard cine images [44]. Sandino et al. developed another model, called "DL-ESPIRIT", that was able to combine the CNN reconstruction with standard reconstruction and provide similar results to standard CS images, accelerating the time of acquisition by twelve times [45]. Kustner et al. developed another approach, a 4D DL-based reconstruction algorithm for 3D Cartesian cine data that was called CINENet [46]. In particular, the CINENet algorithm was able to acquire the left ventricular volume in 15 s, providing the same result as cine images [46].

Image Segmentation
The delineation of the contours of the heart chambers and myocardium (segmentation) is the first step in the processing of CMR images (Table 2 and Figure 7) [50], as the quantitative parameters of left ventricular end-diastolic volume (LVEDV), left ventricular end-systolic volume (LVESV), right ventricular end diastolic volume (RVEDV), right ventricular end-systolic volume (RVESV), and ejection fraction (EF) are derived from it and it has a role in prognostication [51,52].  [53,54]. Automated segmentation methods based on ML and DL have been developed, but occasionally, manual correction is still necessary. As seen in the works on the use of artificial intelligence in the segmentation process, is more difficult to delineate the contours of the right ventricle than those of the left ventricle because of the right ventricle's smaller wall thickness and irregular shape, with the presence of greater trabeculae [62][63][64].
Bai et al. showed that an automatic method based on the use of DL in the process of segmentation and measurement of quantitative parameters in CMR imaging presented a performance equal to that of human experience [55]. Similar results were shown by Penso et al., who demonstrated a good correlation between volume calculated with DL and that calculated with a manual approach [56]. Atrial segmentation could be useful for management of atrial fibrillation, in particular for the planning of atrial fibrillation ablation both in the preoperative period and in follow-up. Xiong et al. developed a CNN on 3D LGE MRI to automatically segment the left atrium [57]. Using DL and manual segmentation, Yang et al. demonstrated the ability of atrial scar segmentation [58]. In parallel, it is well known that the right atrial (RA) area predicts mortality in patients with pulmonary hypertension and is recommended by the European Society of Cardiology/European Respiratory Society pulmonary hypertension guidelines. Importantly, the advent of deep learning may allow more reliable measurement of RA areas in order to improve clinical assessments [65]. Xu et al. proposed a CNN-based automatic segmentation method in noncontrast cine MR images of myocardial infarction areas; this method obtained high consistency with human experience and LGE images [61].

Myocardial Tissue Characterization
ML methods have been applied for the automatic quantification of LGE, overcoming the limits related to its irregular and multifocal appearance, the variation of gadolinium kinetics, and the variability among different centers in regard to accuracy and reproducibility (Table 3 and Figure 8) [59,60,66,67]. ML has also been applied to cardiac relaxometry, specifically T1 mapping, which is useful for identifying changes in extracellular volume [68][69][70]. Radiomics, an ML technique capable of handling high-dimensional data [71], refers to analysis of medical images aimed at obtaining quantitative information, through appropriate mathematical methods and the use of computers, that cannot be detected by simple visual observation by an operator. Various features can be extracted from images, the most important being intensity-and texture-based morphological features (summarized in the term "texture analysis" (TA)) [72]. TA refers to any geometric and/or repetitive arrangement of grey levels and allows the segmentation, analysis, and classification of medical images according to the underlying tissue structure, thus offering the possibility to overcome the limitations of pure interpretation of visual images [73,74]. There are several applications of radiomics and TA: to detect myocardial fibrosis in patients with hypertrophic cardiomyopathy (HCM), to differentiate healthy controls from patients with cardiomyopathy (HCM, amyloid), and to perform scar segmentation in myocardial infarction (MI) in the differential diagnosis between acute and chronic infarction [75][76][77][78][79]. In regard to the application of TA in T1-T2 mapping, Neisius et al. applied it to discriminate between HCM and hypertensive heart disease, while Baessler et al. demonstrated its diagnostic accuracy in both acute infarct-like myocarditis and chronic myocardial inflammation/myocarditis [80,81].

Diagnosis
ML can help physicians in the accurate and early image-based diagnosis of cardiovascular disease. Several papers have used conventional imaging indices as input data to diagnose various heart diseases such as HCM, DCM, MI, and ARVC (Table 4).

Myocardial Infarction
Machine learning methods offer simplification and acceleration of the diagnostic path of MI and are useful as a guide for treatment strategies [88]. Baessler

Cardiomyopathies
Machine learning is an excellent method that allows for the differentiation of various cardiomyopathies [88]. Gopalakrishnan et al. used CMR parameters of the left ventricle, right ventricle, and overall heart from a cohort of 83 pediatric subjects in order to characterize five different cardiomyopathies: HCM, DCM, ARVC, left ventricle noncompaction, and myocarditis [86]. Khened et al. and Wolterink et al. used conventional CMR indices as input for the classification of subjects into five categories (healthy, HCM, DCM, ARVC, and MI) and obtained accuracies of 0.96 and 0.86, respectively, on the test set [62,63,83,84]. Focusing on radiomic features, Nelsius et al. demonstrated an ML model for differentiating HCM from hypertensive heart disease using radiomic features derived from T1 mapping sequences [80].

Heart Failure
Heart failure (HF) is a clinical syndrome due to a structural and/or functional abnormality of the heart that results in elevated intracardiac pressures and/or inadequate cardiac output at rest and/or during exercise. Because of the progressive aging of the general population, heart failure has assumed an increasingly relevant epidemiological problem and currently represents the cardiovascular disease with the greatest increase in incidence. Baessler et al. applied texture analysis on cardiac MRI T1 and T2 mapping and obtained quantitative imaging parameters for the diagnosis of acute or chronic heart failure [82]. Moreno et al. used a dataset that contained cine-MRI images (HF with infarction having an EF < 40%, HF without infarction having an EF < 40%, hypertrophy having a normal EF > 55%, and normal patients having EF > 55%) to describe and characterize heart motion patterns along the cardiac cycle through SVM and RF [89].

Abnormal Wall Motion
Machine learning facilitates postprocessing and image analysis for the characterization of wall motion [88]. Mantilla et al. and Afshin et al. proposed an ML method for classifying normal/abnormal wall motion in left ventricular function using CMR [87,90].

Prognosis
Cardiac magnetic resonance has demonstrated the ability to prognosticate various cardiovascular conditions [6,[91][92][93][94][95][96][97][98]. Cheng et al. showed that greater heterogenicity of LGE in patients with HCM and systolic dysfunction was associated with adverse events [99]. Through the use of biomarkers derived from imaging, ML methods may add value beyond traditional risk scores in the prediction of adverse cardiovascular events, as proposed by the Framingham Heart Study and other cohorts. In the MESA (Multiethnic Atherosclerosis Study) study, there was better prediction of risk, with greater accuracy, in the prediction of cardiac events. [100]. In addition, ML may predict cardiac arrhythmias in patients who have survived heart attacks [101] and therefore have a discriminating power in risk stratification, similar to that of EF and scar size, using the characteristics, position, and size of the scar derived from CMR. It may also be useful in predicting a favorable response to cardiac resynchronization therapy using ECG, clinical, and heart motion analysis data [102].

Limitations
Although machine learning methods have proven effective and have been validated, there are practical difficulties in implementing them in the clinical setting for several reasons, including data quality. Homogenization of clinical data and imaging protocols from different centers is an essential element to check before using datasets as input into a standard ML model. In addition, since clinical data add critical information to images, it is essential to integrate clinical data in electronic medical records to the imaging dataset. Cost is another critical issue, as substantial investment would be required to develop more complex ML methods. Although DL has shown promising results for image processing in the context of high-dimensional datasets and supported by the possibility of real-time reconstruction, in the use of powerful computational algorithms, additional limitations must be recognized. The lack of large, publicly available CMR datasets that can be used to objectively compare different algorithms; the lack of generalization capabilities when previously unseen samples are presented; the difficulties with rare diseases (and, conversely, rare presentations of common entities such as congenital heart disease); and the "black-box" nature of DL algorithms, in that it is often unclear what information is used to arrive at a particular classification or result, represent just a few of the real limitations of this intriguing technique.

Future Perspectives
In this review, we report the use of ML in all aspects of CMR, from the acquisition and reconstruction of images to the derivation of prognostic information. Despite the significant advances described above, there is a need for controlled prospective clinical trials in which ML techniques are compared with human evaluation of magnetic resonance datasets to establish the effectiveness of such methods in clinical practice. The algorithms must be validated using other cohorts besides that under consideration, including data coming from different centers and from different acquisition devices. Furthermore, to compare the performance of different algorithms, it would be important to evaluate the efficiency of each algorithm on a common dataset. Finally, to enable progress in this field, publicly accessible software would be required to allow other groups to run, study, modify, and thus improve software and monitor potential bias in the algorithm [103]. An inevitable and widespread concern is data privacy, so it would be necessary to build a privacy protection algorithm that combines encryption and AI methods in order to obtain fast, secure, and generalizable models. As larger datasets become available, predictive models for both disease progression and simulated response to therapy are expected to develop.

Conclusions
The advancement of techniques in the field of AI can be attributed both to the fact that AI is able to manage high-dimensional data, integrating information from multiple fields, and to the increasing availability of data through mobile applications and data transformation from the healthcare system into digital form. However, the applicability of these approaches in cardiovascular applications remains limited because of the intrinsic peculiarities of cardiac imaging, studies based on small samples that involve a risk of overfitting, and the lack of interoperability of many systems used. However, the use of AI in cardiovascular medicine could pave the way towards better diagnosis and precision medicine and revolutionize the monitoring and treatment of individual diseases.

Conflicts of Interest:
The authors declare no conflict of interest.