Current protocols for stress perfusion CMR involve administering a vasodilator, such as adenosine or regadenoson, to induce vasodilation of the coronary microvasculature and a double injection of a gadolinium-based contrast agent to measure perfusion both at stress and at rest (4). While generally safe, vasodilators carry the risk of severe hypotension, and some patients may require a second injection of a counteracting drug. Injecting a gadolinium chelate twice suffers from perfusion measurement bias for the second injection and from higher than usual gadolinium dosing. Recent research advances have proposed alternatives to both these shortcomings: altered inspired gas concentrations in place of adenosine or other pharmacological vasodilator (5), and use of a blood-pool contrast agent for single-injection assessment of microvascular reactivity (6) (Figure 2).

Inspiring elevated levels of carbon dioxide (CO₂) is well-known for its vasoactive effects on the microcirculation. When inspired CO₂ is elevated, a state known as hypercapnia, different vascular responses are seen, depending on the CO₂ level and the tissue bed. At low levels of elevated CO₂ (<5% inspired CO₂) blood vessels in most tissue beds dilate, while higher levels evoke vasoconstriction in many organs (7). Compared to pharmacological agents, hypercapnia is safer, and its effects can be quickly reversed by restoring normoxia. Furthermore, a very reproducible response can be elicited. However, it was not until the introduction of a device for controlled gas delivery, the RespirAct™ system (Thornhill Research, Toronto, Canada), that hypercapnia was even a practical or reliable vasoactive stimulus for clinical use (5). This system is Health Canada approved and works by adjusting inhaled gas levels to target blood gas levels independent of minute ventilation or breathing pattern (i.e., little subject cooperation is required), thus making blood gas changes a robust, repeatable intervention. A large study of 434 exams in patients aged 9–88 confirmed the safety of this method (8). Most clinical research using the RespirAct has focused on neurological deficits, but some myocardial studies have emerged recently (9, 10). Perfusion increases in the myocardium from hypercapnia is comparable to that from pharmacological agents, as seen in Figure 3A in the healthy human heart, where hypercapnia induced a 3.5-fold increase and adenosine induced a 4.2-fold increase (9). Aside from targeting blood CO₂ levels, the partial pressure of oxygen (pO₂) can also be targeted to bring about hypoxic or hyperoxic effects on blood vessels, further bolstering one’s ability to assess both vasodilatory and vasoconstrictive response.

To detect inducible ischemia, where rest perfusion is normal, a stressor in the form of exercise or a vasoactive agent is needed to evaluate if certain parts of the myocardium are unable to increase perfusion adequately relative to healthy myocardium. The gold-standard for stress perfusion CMR is to inject a gadolinium chelate under maximally induced hyperemia and to measure perfusion during the first pass of the agent (11). The rest perfusion CMR can be acquired either before (4) or after (11) the vasodilator is administered. If the rest perfusion CMR is acquired after the stress perfusion CMR, a short interval after the vasodilator is administered [e.g., 10 min for adenosine (11)] must elapse before a second injection of gadolinium is given and first-pass perfusion again measured. A comparison of the stress perfusion and rest perfusion scans, even in a qualitative manner, then provides a clear depiction of areas affected by compromised vasoreactivity.

There are a few shortcomings with the gold-standard cardiac stress CMR described above. First, there is no room for error when it comes to contrast injection, and since there is only one window of opportunity to capture the first passage, the acquisition sequence and the contrast injection must be timed perfectly. Second, since the interval between the two injections is short relative to the elimination half-life of gadolinium chelate, an elevated baseline signal for the second injection will inevitably introduce errors in quantitative perfusion estimation. Even if T1 mapping is performed before each injection to convert signal intensity to gadolinium concentration, the pharmacokinetics is altered for the second injection due to residual contrast from the first. A final consideration is that the total contrast dose is double that of typical contrast-enhanced MRI exams.

Recently, the concept of blood-pool imaging of vasoreactivity was proposed to circumvent the limitations of conventional stress tests (6). A T1-reducing agent is introduced to sensitize T1 relaxation times to the intravascular compartment, as native T1 mapping is insensitive to blood volume changes from vasomodulation (7). A low dose of a gadolinium-based blood-pool agent, Ablavar^® (gadofosveset), is introduced intravenously and allowed to reach a quasi-steady state concentration after 10 min. This steady state lasts approximately 40 min in rats, even longer in humans due to a slower elimination rate. Over this time interval, one can administer a vasoactive substance, such as adenosine or hypercapnia, or even multiple stimuli in succession, and measure how much the microvascular blood volume changes. It is noteworthy that blood-pool imaging of vasoreactivity was initially conceived to permit a more specific and direct measure of vasomodulation (6). By looking at changes in blood volume, confounding factors—such as heart rate variations—implicit in conventional perfusion imaging are eliminated, thereby yielding a specific measurement of vasoreactivity. This approach also eliminates quantification errors resulting from double contrast injections. Several groups have since adopted the blood-pool imaging approach and applied it to myocardial stress testing (12, 13). Instead of Ablavar^®, however, whose production was discontinued in 2017 due to low sales, these papers used Feraheme^® (ferumoxytol), an iron-based compound currently still approved in the U.S. as an iron supplement and used off-label as a blood-pool MRI contrast agent. Figure 3B illustrates blood-pool MRI of vasoreactivity in the kidneys, where blood-pool vasoreactivity imaging was first demonstrated, and the heart.

Blood-pool MRI contrast agents are valuable beyond their intended utility for angiography. They enable lower dosing and a much longer window for image acquisition. They also provide an alternative approach to assessing microvascular reactivity that permits higher specificity, quantitative accuracy, and spatial resolution than is possible with conventional double first-pass perfusion assessment of MVD. However, at the time of this writing, there is no clinically approved MRI blood-pool contrast agent on the market. Perhaps when more clinical indications in which blood-pool agents provide a clear advantage come to light, or when new gadolinium-free versions become accessible, there will be an incentive for pharmaceuticals to bring MRI blood-pool agents back into the clinic.

In 2007, the first type I collagen-targeted MRI contrast agent, EP-3533, was introduced by Caravan et al. (16). The agent was a cyclic peptide with three gadolinium moieties, and the dense scar tissue of a mouse myocardial infarct was shown to enhance preferentially relative to surrounding healthy myocardium. Since the first demonstration in an infarction model, subsequent studies extended its application to mouse models of liver (17), pulmonary (18), and skeletal muscle (19) fibrosis. All studies confirmed superior sensitivity and specificity of collagen-targeted agents for fibrotic tissue compared to non-targeted counterparts. However, EP-3533 contained a linear gadolinium chelator used in DTPA, thus raising the risk of gadolinium toxicity. Recently, the collagen-targeting peptide in EP-3533 was attached to a more stable gadolinium macrocyclic chelate (20). Theoretically and in practice, macrocyclics are more stable than linear ones, and fewer reports of gadolinium deposition are associated with macrocyclic chelates. Nonetheless, because agents with an affinity to collagen, or any targeted agent, are designed to have longer residence time in the body than would a non-targeted agent, the safety profile must be conclusive before translation into humans is approved. Despite the fact that EP-3533 is currently not approved for clinical use, however, if its clinical translation is eventually successful, there would be a profound impact on disease detection, staging, and treatment monitoring.

Non-gadolinium agents targeted to collagen have been reported, but these are relatively few compared to EP-3533. For example, c(RGDyC)-USPIO is an iron oxide nanoparticle that targets liver fibrogenesis by imaging activated hepatic stellate cells that specifically engulf the contrast agent (21). Another example are manganese porphyrin-based contrast agents that have been used for labeling and tracking collagen scaffolds (22, 23) in tissue-engineering applications. These manganese agents offer very high sensitivity, and although they bound through non-specific mechanisms, it is possible to introduce new conjugation methods to specifically target collagen. The advantage of utilizing manganese-porphyrin is that manganese, unlike gadolinium, is a natural substance and is needed in small quantities by the body. Furthermore, the porphyrin ring provides a thermodynamically stable product such that demetalation in vivo is highly unlikely, even in long retention applications such as fibrosis imaging. Development of manganese porphyrin-based fibrosis contrast agents is currently geared toward both young scar (fibrogenesis) and old scar (fibrosis).

Cardiac MRI has also developed its own version of strain imaging. Several strain CMR techniques currently exist, but none is fully developed and standardization is a distant goal (27). Strain CMR was first introduced by Zerhouni et al. in which a magnetic label in the form of black lines was superimposed at the beginning of a dynamic CINE sequence and their deformation tracked through the cardiac cycle (28). This first strain imaging method is known as CMR tagging and is akin to speckle tracking in strain echocardiography (29). Both short- and long-axis images are acquired to measure motion in all three directions, which requires a longer series of breath-holds and is susceptible to image mis-registration errors. Furthermore, spatial and temporal resolutions are low and acquisition times long. Later methods for strain imaging were introduced to improve upon the capabilities of CMR tagging. Phase velocity mapping allows calculations of myocardial velocities in three directions and provides higher spatial resolution, albeit at the cost of lower temporal resolution (30). Displacement encoding with stimulated echoes (DENSE) encodes displacement in the phase image (31); while this technique inherently has low signal-to-noise (SNR), acquisition times are short. Finally, strain-encoded (SENC) imaging is similar to CMR tagging in that magnetization tags are used (32). However, rather than placing tags orthogonal to the imaging plane, SENC places tags parallel to the imaging plane, and only short-axis images are needed to measure longitudinal strain. Post-processing for SENC is quick, but one downside is that radial strain cannot be measured. Of all these strain CMR techniques, only CMR tagging has seen the most validation (33).

Feature tracking, unlike the methods described above, is a strictly post-processing technique and is applicable to any modality, not only MRI. No additional imaging is necessary, and standard CINE images can be retrospectively analyzed using feature tracking of cavity/myocardial borders to calculate strain. It is already available clinically and has demonstrated both reproducibility (34) and accurate delineation of infarct territories (35). However, as a contour-based tracking method, feature tracking cannot assess mid-wall strain and is limited by low spatial and temporal resolution, poorer performance on regional strain imaging, and lack of information on rotational strain (36). Therefore, despite its ease of use, feature tracking CMR tagging remains the accepted gold standard imaging modality for strain quantification.

Despite the lack of standardization for strain CMR, myocardial strain imaging has been investigated in different cardiac pathologies. Perhaps the most common application is ischemic heart disease. After a myocardial infarction, myocardial strain components are most impaired in the infarct (35) and strain analysis can identify which tissue segments will recover (37). However, given the availability of late gadolinium enhancement to identify infarct zones robustly, the role of strain CMR in this pathology is at best confirmative and not complementary.

A more difficult phenotype to diagnose is non-ischemic heart disease, where late gadolinium enhancement typically does not reveal evident tissue abnormalities. In this setting, strain CMR can be invaluable. Strain measurements have shown impairment in patients relative to healthy control subjects. For example, in hypertrophic cardiomyopathy, mid-wall fibrosis underlies abnormal systolic strain measurements with or without late gadolinium enhancement (38). Myocarditis also presents impaired myocardial strain in the absence of any other abnormality (39). Perhaps most notable is the potential for strain CMR to provide early indication of disease development. It has been demonstrated that abnormal alterations in strain parameters manifest much earlier than reductions in ejection fractions, hypertrophy, or even presentation of symptoms (40). It is for this reason that strain imaging is routinely recommended in patients undergoing chemotherapy, who are at risk for cardiotoxicity from chemical agents like doxorubicin.

Strain echocardiography is a validated technique for assessing subtle myocardial changes. However, it is used strictly for larger muscular structures. It is unsuitable for thin structures of the right ventricle and the two atria. Strain CMR, on the other hand, is the gold-standard for assessment of the right ventricle and is the first modality to demonstrate right ventricular strain analysis (41). Diseases of the right ventricle, while less common than those of the left ventricle, affect a significant patient population: those living with congenital heart disease, pulmonary hypertension, or arrhythmogenic right ventricular cardiomyopathy. Similarly, atrial strain measurement is only possible on CMR. Essentially, deformation of the atrial walls can become aberrant in diseases characterized by high filling pressures and diastolic dysfunction. Pathologies exhibiting these properties include HFpEF (42) and myocarditis (43), where atrial strain CMR have demonstrated promising results. Given the utility of strain CMR in characterizing not only pathologies that strain echocardiography can detect but also those that echocardiography cannot, a very real need exists to bring about consensus amongst MRI vendors to standardize strain CMR and ensure quality control. Figure 5 illustrates a few potential applications for strain CMR.

13P MR spectroscopy assesses muscle energy metabolism by monitoring the ratio of phosphocreatine (PCr) to ATP. Cardiac muscle is like any other muscle in our body in that it requires a substantial supply of ATP to power contractions. When an ATP molecule is hydrolyzed to release energy, it produces adenosine diphosphate (ADP) and an inorganic phosphate as by-products. To maintain a constant energy reservoir, ATP must be regenerated; the molecule PCr is required to transfer high-energy phosphate to convert ADP into ATP. During a period of low energy requirement, excess ATP is used to convert creatine back to PCr. Therefore, PCr acts as a high-energy reserve to maintain normal ATP levels regardless of fluxes in energy demands. Since ATP levels are relatively stable, assessment of the muscle cell’s energy store is achieved by monitoring the ratio of PCr and the three phosphate groups on ATP. Any decrease in the PCr/ATP ratio is indicative of a depleting PCr store.

Heart failure is a prime example of a disease phenotype where 31P MR spectroscopy is informative. The healthy heart forms ATP by oxidizing both fats and glucose, and the proportion from each substrate is adapted to the environment—fatty acids are the major substrate in the healthy and trained heart, whereas glucose utilization increases in heart failure. Regardless, utilization of both substrates is reduced in the failing heart, resulting in a decline in the PCr/ATP ratio (45). This observation has been confirmed in both human studies of advanced heart failure [e.g., dilated cardiomyopathy (46)] and induced heart failure in mice (47). Patients with a confirmed HFpEF diagnosis also present a lower PCr/ATP ratio, approximately a 25% decrease, compared to healthy control subjects (48). Perhaps even more illuminating is that the comorbidities associated with HFpEF—such as diabetes (49) or the typical hypertrophic phenotype of HFpEF (50)—present similar reductions in the PCr/ATP ratio. The discriminatory power of imaging energetics extends to subclinical cardiac dysfunction—obese patients demonstrate a modest but significant decrease in PCr/ATP at rest, a decrease that is further exacerbated during a dobutamine stress test (51). These examples illustrate that clinical implementation of cardiac 13P MR spectroscopy, while complicated, is not impossible.

The most significant limitation of 31P MR spectroscopy is the low sensitivity of the 31P nucleus. The gyromagnetic ratio of 31P, at 17.2 MHz/Tesla, is 40% that of the hydrogen proton. Furthermore, phosphate metabolites exist at relatively low concentrations in the body. Therefore, to attain adequate SNR, voxels tend to be large and acquisitions prohibitively long for clinical implementation. In recent years, increased use of higher field magnets has translated to higher SNR, which can be traded for shorter acquisitions or better spatial resolution. More significantly, several acceleration methods have been proposed for 31P MR spectroscopy. Echo-planar spectroscopic imaging, which exploits rapidly oscillating readout gradients to encode both spectral and spatial information, was demonstrated for 31P MR spectroscopy in the human calf muscle with a scan time of 10 min (52). A spiral version of the oscillating gradient has also been demonstrated for muscle 31P MR spectroscopy (53). Despite a much faster scan time, however, both techniques incurred an SNR penalty. Perhaps more interesting is a subspace technique called SPICE (for SPectroscopic Imaging by exploiting spatiospectral CorrElation) introduced in 2014 for rapid, high resolution proton spectroscopy (54). The same group later applied SPICE to 31P MR spectroscopy, wherein they achieved a resolution of 6.9 × 6.9 × 10 mm³ and 15 min scan time when imaging human calf muscle on a 3 Tesla scanner, with the spatial and temporal resolution improving to 1.5 × 1.5 × 1.6 mm³ and 30 frames/s, respectively, at 9.4 Tesla (55).

Other challenges also hamper clinical translation. 31P MR spectroscopy is inherently complex, but obtaining a good 31P spectrum of the heart is even less trivial, even if on a qualitative level. Both cardiac and respiratory motion complicate acquisition, and acquisition time is prolonged due to the need for triggering. Obtaining a homogenous B1 field is also important but is time-consuming and often difficult. Furthermore, qualitative assessment of PCr/ATP is suboptimal in the context of detecting early disease, as shifts in metabolism may not be yet quite pronounced. A better approach would be quantitative measurement of absolute PCr and ATP levels. While more tedious, one can achieve quantitation by including a reference phantom outside the body. Continued development of cardiac 31P MR spectroscopy toward clinical adoption will need to see advances in high-SNR coil design, rapid imaging, dynamic shimming, and greater access to higher field strength scanners. All of these will improve spectral and spatial resolution and reduce acquisition times. With continued technological advances in these areas, metabolic imaging will one day be part of routine clinical workup for cardiac patients. Figure 6 illustrates a 31P MR spectroscopy experiment, existing challenges to clinical translation, and applications that may benefit greatly from this technology.

The premise for acceleration can be formulated as an optimization problem, where the original high-resolution image must be reconstructed from a sub-Nyquist undersampling of the acquisition domain, namely, k-space. The original strategies would provide aliased images (aliased due to severe undersampling) to a neural network and train against ground-truth unaliased images from fully sampled datasets (60). The neural network is essentially tasked with learning the Fourier transformation to convert images from the image domain to k-space and vice versa, as well as learning features of the training dataset. Variations on this image-to-image approach also exist. For example, one can feed in undersampled k-space datasets and train against fully sampled unaliased images (61), or one can feed in undersampled k-space datasets and train against fully sampled k-space (62). For the latter approach, the neural network learns how to fill in missing k-space datapoints.

Most acceleration applications to date exploit the convolutional neural network architecture (63). This choice is fitting for network applications involving images, because images can be more sparsely represented as a summation of two-dimensional kernels representing edges in different orientations, different shapes, and any “building-block” structure. Different layers of the neural network represent features at different scaling factors. The deeper the layer, the coarser the feature. More recently, a novel architecture has been investigated with great interest. This architecture, known as the transformer (64), differs from the convolutional neural network in that it is based on a concept called attention and is ideally suited to handling sequential tasks while capturing long-range dependencies. Originally conceived for natural language processing, where one needs to look at dependencies between widely separated words to construe the meaning of a sentence, the concept of attention can also be extended to an image, where pixels far apart but contributing to the same structure must be recognized as an intact entity. Applications of the transformer to image denoising, motion deblurring, and defocus deblurring have been reported (65). In the area of image reconstruction alone, transformers have demonstrated superior performance compared to conventional convolution neural networks (66, 67).

Adaptions of the reconstruction techniques described above can involve, as examples, training on both the image domain and k-space or getting the neural network to uncover the optimal undersampling pattern in k-space. The first method, also known as dual-domain reconstruction, makes the reasonable and logical assumption that providing both object and frequency domain data for training should improve reconstruction quality (67, 68). While this is, indeed, the finding, it also comes to no surprise that the improvement in reconstruction quality is modest, since the information contained in k-space is identical, no more and no less, than the information contained in the image. The second approach of identifying the optimal sampling region of k-space invariably results in an oversampled center of k-space and sparser sampling in the peripheral spatial frequency regions (69). This behavior is also expected. Perhaps a useful piece of information that can be gleaned from these studies is the ideal probability distribution of k-space sampling. Most, if not all, of these sampling patterns cannot be realized upon implementation, but they can inform on how to best sample k-space in a statistical manner given real constraints on gradients.

Despite the successes of different neural network architectures in achieving high quality reconstruction from undersampled acquisition data, there are quite a few technological shortcomings that should be recognized. Most obvious is the availability of large datasets for training. There are a handful of MRI databases currently available, such as the UK Biobank, which contains the largest collection to date of acquisitions in various body regions. For instance, it has an abundant repository of different types of brain and cardiac images. Assuming that access to a large database is not an issue, the next hurdle, one much more difficult to overcome, is the high dependency of training on the acquisition itself. If the tissue type changes, or any of the acquisition parameters (e.g., imaging plane orientation, spatial resolution, image contrast, field-of-view, sequence type, and vendor type) changes, the network must be re-trained for the new scenario. It is easy to appreciate that this limitation can severely curtail the clinical utility of deep learning-assisted acceleration.

Recently, researchers have proposed so-called physics-based deep learning methods to partially address some of the limitations described above. These are dubbed physics-based, because the neural network is trained to solve a supervised learning task and, at the same time, adhere to any physical laws governing the system of interest. As an example, rather than treating each image as a new reconstruction problem, we can exploit known information in regard to the anatomy of interest or aliasing artifacts specific to different k-space trajectories. In the first physics-based reconstructions, the undersampling pattern and coil sensitivities, both with known effects on the reconstructed image, were taken into account to solve an inverse problem (59, 70). However, these applications employed fully sampled data as a reference during training. In cardiac imaging or myocardial perfusions scans, where it may be infeasible to have fully sampled training datasets, one is compelled to train on undersampled data. In 2021, Yaman et al. (71) introduced a self-supervised deep learning approach to train physics-guided reconstruction without fully sampled reference data. While their platform was demonstrated on brain and knee data, it would be instructive to extend their network to cardiac MRI. Figure 7 presents a schematic of a deep learning-assisted reconstruction of undersampled MRI data and existing challenges that must be overcome to permit widespread clinical adoption.