Following detailed history, clinical examination, electrocardiogram (ECG) and laboratory tests that may elucidate features of a specific underlying etiology or secondary organ dysfunction, imaging techniques play a crucial role in confirming the diagnosis, ruling out other competing causes for LV dysfunction, further evaluation of the etiology and in guiding treatment strategies.

Whilst two-dimensional echocardiography is often first line in the diagnostic imaging pathway and has an additional role in both early and follow up function assessment in DCM patients, its role in defining an underlying etiology is limited, particularly with the compromise that occurs in light of inadequate acoustic windows and poor endocardial border definition. Furthermore, due to the inherent geometric assumptions that perform well in healthy individuals with normal sized hearts, volume assessment in those with distorted ventricular size and shape is less reliable, with significant intra- and interobserver variability.

CMR is well-placed in this respect, with unrestricted field of view and high spatial resolution to capture global and regional changes in structure and function irrespective of ventricular geometry or patient habitus (14). As there is less operator dependence for endocardial delineation, the interobserver reproducibility variability for volume and EF quantification is less for CMR than it is in echocardiography (14). This is ideal for both the initial evaluation, where decisions on initiation of medical therapy are based on LVEF thresholds, but also to carefully monitor progression of the disease and the appropriate selection of those who require device implantation. The integration of perfusion and whole heart angiography enables the exclusion of significant coronary disease with a high accuracy, thereby reducing the need for separate ischaemia assessment by computed tomography (CT) or invasive coronary angiography in the initial work up of DCM (15). Thus far, routine use of CMR for diagnosis alone has not been shown to significantly improve the clinical identification of non-ischaemic heart failure causes (16). However, complementary information is offered with tissue characterization and parametric mapping sequences that enable assessment of changes to intrinsic myocardial properties correlating with altered biological pathways. These additional features offer the potential to aid the differentiation of the underlying etiology, enable prognostic assessments and guide treatment options. Although there remains a lack of data from large randomized controlled trials asserting the role of contemporary CMR on impacting patient outcomes, the evolving landscape of techniques and applications for in-depth phenotyping paired with advanced analytics pave an important path toward CMR-guided precision care in the DCM population.

Standard CMR provides the gold standard for biventricular volume assessment, further allowing for accurate documentation of systolic function, which is imperative for the investigation of all comers with heart failure. Even though current clinical practice focusses on these static measures obtained from only two end time points of the cardiac cycle, due to real time acquisition over multiple phases, cine-CMR possesses additional information on dynamic volumetric changes. Consequently, it is feasible to generate volume/time profile curves that allow evaluation of continuous ventricular volume changes and extraction of more sensitive parameters of cardiac function such as peak filling rates (see Figure 2) (17). From this, additional indices of filling and ejection are possible to obtain simultaneously, with the potential for more detailed analysis of both systolic and diastolic function (18). Differing LV filling patterns have already been suggested to exist amongst DCM patients with direct implications on the classification of functional status and predicting adverse outcomes (19, 20). However, most studies that assessed these parameters in the DCM phenotype were significantly limited in the diversity of structural and functional heterogeneity seen in most contemporary DCM cohorts, thus hindering full exploration into the evolutions of filling and ejection patterns in different subgroups and their varied clinical outcomes (18, 21). Such studies are warranted but the current tools for obtaining these parameters are limited by the extent of user-interface involved in semiautomatic processing, whereby contours are determined not only at each slice level but also at each time point or phase prior to the computation of volumes needed to produce the curves for each patient.

Unique tissue characterization sequences add a further dimension to the investigative prowess of CMR in the evaluation of the DCM phenotype. The ability to non-invasively assess and quantitate myocardial tissue properties makes CMR well-suited to unravel the onset and extent of pathogenic processes occurring within the myocardium, that could only previously be determined through high-risk invasive cardiac biopsy.

CMR can confirm and reproduce the assessment of LV mass, volumes, and LVEF, all of which are important indicators for a worse prognosis in severe DCM and other causes of heart failure; the latter two markers being key targets for reverse remodeling and myocardial recovery (56–59). The main limitation of these measures for predictive outcomes is that they are often assessed at initial evaluation, failing to account for the dynamic nature of the disease with favorable response to therapy for a significant proportion of patients; concurrently, they are less sensitive for those with mild-moderate dysfunction who are still prone to significant risk of sudden cardiac death (25, 26, 60).

Risk stratification in this setting is difficult and the current focus of this has shifted toward a multiparametric, dynamic approach, which attempts to incorporate potential biomarkers from biochemistry, ECG signals and imaging (12).

There is increasing evidence for applications within CMR to guide prognostication and subsequent clinical management in DCM. Whilst the majority of these applications for risk prediction are captured through routine assessment, the additional tools, and longitudinal follow-up capability is still regarded as an investigational field of interest within the setting of CMR (14). These current and potential clinical CMR applications in the risk assessment of DCM are outlined in Table 1.

Much of the current CMR tools for characterization and predicting outcomes in DCM rely on multiple dedicated imaging sequences, followed by significant time devoted to qualitative post-processing in the evaluation of structure, function and tissue characterization. Despite their feasibility and utility, they are often not fully exploited in clinical practice due to these time constraints on clinical workflow. Even if employed, this often occurs ad-hoc and limited to one or two additional parameters evaluated in uniform manner, rather than assimilated in multi-parametric fashion for personalized characterization and risk stratification. Fully integrated analysis of all these features and metrics could aid better selection of patients who might benefit from earlier medical intervention, need closer surveillance regardless of LVEF, and those who we can more confidently discharge or halt medical therapies following improvement in their cardiac status (5, 13, 66).

It is increasingly appreciated that DCM has a genetic basis, with disease causing variants identified in up to 40% of families of DCM and 25% of presumed sporadic cases (67). Some of these genetic mutations can predispose carriers toward significant brady- or tachy-arrhythmias, or in the presence of environmental factors such as alcohol, can be the driver for a more severe phenotype, and there is a suggestion that a genetic basis could explain the higher prevalence of DCM seen in particular ethnic groups (67–69). There has been an expansion in the reported breadth of genes associated with the DCM phenotype, particularly in recent times with the arrival of next generation sequencing methodology (70, 71). However, robust genotype-phenotype correlations are not always feasible as the genes implicated encode proteins with a variety of different functional properties, making it challenging to harmonize the extent of genetic influence on the spectrum of structural and functional changes in those with DCM (71). Furthermore, it is challenging to clinically define and manage the large number of variants of uncertain significance (VUS), inadvertently arising as a result of the high throughput of current genetic testing.

Being able to discern the full scope of genetic influence in those with DCM will further help tease underlying drivers of disease manifestation and offer the opportunity to establish a deeper characterization of the phenotype. Coupled with the challenge to examine genetic influence on the spectrum of structural and functional changes, the addition of mutation status to clinical and imaging parameters may improve risk stratification and potential treatment strategies beyond the consensus management for heart failure with reduced ejection fraction (HFrEF) (72).

The opening act to CMR characterization involves image acquisition and segmentation, prior to feature extraction. The quality of this step is essential to the outcome of further downstream analysis and provides baseline cardiac parameters as well as the untapped features that might better describe cardiac function in specific cohorts. Assessment of the anatomical features following cine acquisition includes assessment of myocardium, pericardium, all 4 cardiac chambers, valves and vascular connections. Typically, a visual quality assessment is needed first to ensure the signal-to-noise ratio is enhanced by adequate positioning and breath-holding technique, limited blurring by cardiac gating, and appropriate planning for each subsequent image plane acquisition. Following the anatomical review, long and short axis cines are acquired that enable dynamic views of the global heart function. Segmentation by manual planimetry, or by semi-automated methods with clinician oversight enables the reproducible 3-dimensional (3D) assessment of atrial and ventricular volumes, LV mass and EF quantification.

Studies have shown that DL methods can outperform conventional ML, and in some cases, even better in both detection and segmentation tasks analyzed by human expertise (80, 81). CNNs are the technique of choice and the most successful type of models for image analysis (82, 83). Efficacy of the DL models is often assessed in the form of pixel classification accuracy. Although different methods for assessing this exist, the preferred evaluation metric for DL-based segmentation approaches is the Dice metric, which evaluates the overlap between automated segmentation and the ground truth segmentation. The Dice metric has values between 0 and 100%, where 0 denotes no overlap and 100% denotes perfect agreement.

One of the main challenges to implementation of CNNs in medical imaging is the lack of high-quality expert annotated data, available for training the DL network. Furthermore, these datasets often suffer from class imbalances due to certain conditions being encountered less frequently, thereby making it more difficult for a CNN to generalize and limiting large scale CMR evaluations. As highlighted in Table 2, whilst the segmentation performance of state-of-the-art DL methods is commended, it is evident that the number of DCM cases encountered in these evaluations has significantly low representation. Given the heterogeneity of this condition with many individuals at presentation subject to highly remodeled ventricles and rotated cardiac axes, these current automated segmentation methods may not yet be robust enough for the deployment and evaluation of this phenotype.

To ensure any of these methods can translate into clinically useful tools in the evaluation of DCM, it is essential they are complemented by high quality datasets, that help improve the accuracy of segmentation and classification tasks, whilst providing large variability in terms of the clinical phenotype and image acquisition modules, thus enabling generalizability (87, 91). However, manual annotations of large datasets that are able to encompass this scale of heterogeneity is no easy feat, being costly and requiring extensive expert time for good quality annotation. This could be partly overcome with data sharing initiatives and collaborations between CMR centers to obtain large repositories of images with associated clinical information. This is not inevitably a seamless solution, as there are often ethical and legal requirements to satisfy within all participating sites, with limits set on how and where specific data can be utilized during the development and deployment of the pipeline.

Encouragingly, over the years open technical challenges and several publicly available datasets have been made available, helping to unravel this generalizability issue (93, 94). The UK Biobank (UKBB), although limited to a single CMR vendor, provides one of the largest imaging datasets facilitating the exploration of DL capabilities in a large general population whilst solving issues relating to ethics and clinical data aggregation. This was harnessed recently in a genome-wide association study of CMR-derived LV measurements in ~36,000 participants from the UKBB to study the relationship between genetic variants associated with LV structure and function, and risk of incident DCM (95).

Another way of improving the generalizability during training and take advantage of the limited amount of high-quality labeled data is the strategy of data augmentation. It is possible to artificially increase the variation of examples encountered by applying random transformations such as image rotation by certain degrees, image scaling to increase variations in organ size, changing image orientation with random horizontal or vertical flips and even inclusion of random “noise” to images (94, 96, 97). Whilst this option effectively enables the acquisition of more labeled data, the diversity in practice may still be limited in terms of reflecting the full spectrum of the DCM phenotype and the pixel-level differences of images obtained from different CMR vendors. The breakthrough in improving this generalization of networks to reliably segment heterogenous phenotypes acquired from different CMR vendors and clinical sites was demonstrated recently by Chen et al. (94). Unlike the efforts to solely re-train or fine tune networks to improve the performance on a specific dataset, they explored the pre-processing step of data normalization enabling their network to deal with the distribution changes amongst input features from multi-source images. This overcomes the small differences in features arising from images obtained from different scanners and the overfitting to distribution changes that occurs with network development from a single source. Along with data augmentation strategies, their approach achieved encouraging results in terms of reliable segmentation accuracy across test images from multi-scanner and site domains (mean Dice metric of 0.91 for the left ventricle, 0.81 for the myocardium, and 0.82 for the right ventricle from a single site dataset; and 0.89 for the left ventricle, 0.83 for the myocardium from a multi-site dataset).

Recent DL techniques have enabled rapid expansion in the CMR domain to achieve robust contour identification and accurate classification performance, whilst significantly minimizing the extent of post-processing involved in volumetric data calculations (98). Emerging approaches that have recently become commercially available, have further demonstrated the feasibility and precision of fully automated dynamic measurement of LV volumes (99, 100). Utilizing anatomical localization methods to determine relevant boundaries between structures, contours are created in consecutive frames of the cardiac cycle with LV volume/time curves derived at no extra time expense and with high correlation to the manual reference technique (99). Whilst such applications show promise for the application to evaluating DCM cohorts, the development and training of this technique has been based on very limited representation from such patients, with less accuracy seen in those with significantly dilated ventricles and whose impaired breath-holding technique can lead to significantly more artifacts, reducing image quality. Furthermore, the details of the DL pipeline are not disclosed by the manufacturer and this lack of transparency will make it difficult to optimize the current algorithm in order to generalize to other scanners and more diverse patient cohorts.

Motivated by these limitations, Ruijsink et al. (17) developed a robust, accurate and fully automated framework for CMR cardiac function analysis which included comprehensive quality control detection using a CNN to limit erroneous output. Segmentation of both ventricles in all frames was then executed utilizing a 17-layer 2D-FCNN, prior to an iterative alignment process to correct for any differences in breath-holding and motion. After validating their framework that presented with high correlation to manual analysis, biventricular volume curves were generated for over 2,000 healthy individuals to obtain a more detailed description of cardiac function, inclusive of diastolic parameters such as peak early filling rate, atrial contribution, and peak atrial filling rate. These parameters stratified healthy patients by age categories, with lower filling rates correlating with older age–a relationship consistent with the known increase in ventricular stiffness with age (101). Considering that these LV filling patterns also appear capable of distinguishing the different categories of diastolic dysfunction characterized on echocardiography, it is anticipated that this method could enable within DCM subgroups detection of those with persistent diastolic impairment despite LV systolic recovery on medical therapy, and identify patients with subclinical disease who will require closer surveillance (102). These parameters are feasible with no additional imaging outside routine care and can occur at no extra time-cost whilst the routine clinical analysis is ongoing. In terms of the potential clinical application to evaluation of the DCM phenotype, it is significant that the method employed by Ruijsink et al. (17) performed similarly well in unseen patients with cardiomyopathy as well as those without cardiac disease. It has been reported from studies on emerging 4D flow CMR, that DCM patients have altered and heterogenous diastolic flow patterns that occur due to abnormal filling kinetics and varying degrees of pathological geometrical configuration of the LV (103). This highlights the potential role offered by fully automated LV filling assessment in differentiating those with persistent altered filling patterns and abnormal diastolic flow, thereby remaining at risk of relapse compared to those who have truly achieved recovery and remission. Based on this promising AI tool, current work by this group is also exploring the innovative use of GANs to generate realistic CMR images from any domain in order to advance the generalization of the network and robustly deal with clinical CMR data from multiple centers, vendors, and field strengths (104). Given the feasibility to evaluate both ventricular filling profiles, and the suggested prognostic role of serial revaluation of RV function in the follow up of DCM, the characteristics and clinical utility of RV filling patterns over time will be another area of application in the DCM population (105).

State-of-the-art algorithms utilized in scar segmentation are commonly semi-automated, fixed-model approaches where the pixel intensities of scar regions are exploited through a process of thresholding (106). This requires a user-selected area of interest and knowledge of the nearby intensities of healthy nulled myocardium, prior to operating a region growing process to segment the scar region. These methods are currently popular for segmenting contiguous regions of scar, and are highlighted for their reproducibility, with encouraging performance against consensus segmentation by experts (106). Whilst simple to implement, they remain heavily user dependent for pre-processing with respect to definition of the myocardial borders, activating the boundaries of interest, initialization for region growing in each slice, and the subjective baseline selection of remote healthy myocardium as well as the perceived extent of scar. More automated approaches have been developed to help minimize the degree of user interaction whilst maintaining reproducible performance (106). These methods mostly utilize clustering techniques to fit data of different tissue signal intensities in order to characterize the voxels belonging to scar regions (106). Whilst they show good correlation with the fixed-model approaches in accurately identifying LGE, they unfortunately are not robust enough for clinical translation due to failure to accurately segment scar where CMR-LGE images are affected by noise or share homogenous signal intensity distributions within myocardial boundaries and other nearby tissues (106–109). This limitation is particularly important as most of these traditional methods have been validated on CMR-LGE images obtained from patients with coronary disease, where the pattern of scar is subendocardial as opposed to that seen, if present, in DCM and other non-ischaemic cohorts, occurring in the mid to epicardial wall segments, where tissue intensity homogeneities are more likely to be encountered.

As the attention of CMR segmentation transitions toward more DL-based approaches, it is hoped that these innovative techniques will also facilitate a more practical and reliable means of achieving standardized quantification of LGE. This is highly desirable, given the suggestion that even after adjustment of LVEF, the proportion of LGE assists the clinical stratification of DCM patients who are prone to a higher risk of death and hospitalization (28). Not to mention, the ability to efficiently characterize border zone tissue–areas of variable transition between scar and normal tissue that have arrhythmogenic potential, thereby helping to identify those at risk of malignant arrhythmias and more likely to benefit from ICDs (110).

A successful FCNN architecture, the ENet was recently harnessed to deal with the task of scar segmentation (111). Several variants of this popular architecture have been adapted to enhance the accuracy of different cardiac segmentation tasks (see Table 2). However, the pursuit of scar segmentation is a relatively new concept. In this work by Moccia et al. (111), the ENet was adapted and evaluated to see if pre-identified LV regions could enable more accurate scar segmentation than current methods, and furthermore whether a fully automated output of scar segmentation was feasible and maintained a similar or improved accuracy. As a proof of concept, this method showed both protocols were able to identify scar on the CMR-LGE images without the need for pre-processing extraction steps. However, it was the semi-automated method with a priori knowledge of the restricted myocardial boundary in which to search for scar, that outperformed state of the art CMR-LGE segmentation algorithms and was closest to expert annotation (with a sensitivity of 0.88 and Dice coefficient of 0.71). This is still an important breakthrough, holding advantages over current efforts to quantify scar by minimizing subjective evaluation, user interaction and any parameter tuning prior to implementation. Expanding the training datasets to incorporate the variability of scar seen in those with a sole DCM phenotype and those with accompanying embolic sub-endocardial scar, could be an encouraging start to help encode the high variability of scar dimensions seen in this population. This study provides an important step forward in the clinical practice of scar quantification, and by enhancing the pixel classification through training, the ENet would not only acquire improved segmentation performance, but would be more generalizable to the DCM population.

In conjunction with acquiring diverse DCM datasets, image data augmentation is another common method to artificially boost training datasets in order to improve performance and generalizability of a deep learning model. This may be particularly relevant with regards to the DCM population, where a disconnect exists between high demand for sufficient training images and the variability of scar presence across the spectrum of patients, in essence, limiting the real-world availability of training examples (24). This was recently explored in the simulation of scar tissue on the LGE-CMR images of healthy patients (112). Lau et al. (112) utilized their GAN framework which could additionally incorporate domain-specific knowledge, to simulate various scar tissue shapes in different positions. These images were highly realistic as demonstrated by the improved segmentation prediction of scar tissue pixels correctly identified during testing from 75.9 to 80.5% and the qualitative assessment that imaging experts were unable to reliably distinguish between simulated and authentic scar.

A stream of work has focused on extending the use of a trained FCNN to assist analysis of myocardial tissue characterization by means of automated native T1 mapping (89). With good agreement to manual calculations, this showed promise for an automated pipeline to minimize the workflow involved in quantifying global T1 characteristics. This was validated on a single scanner, with further study needed to see if this method can be applied to other mapping sequences such as the modified Look-Locker inversion recovery (MOLLI) or the contemporary shortened modified Look-Locker inversion (shMOLLI) method, that is more acceptable and compatible with typical limits for end-expiration breath-holding in patients (113). Puyol-Antón et al. (114) evaluated an automated framework for tissue characterization using the shMOLLI method at 1.5 Tesla using a Probabilistic Hierarchical Segmentation (PHiSeg) network. This method models the probability distribution of pixel-wise segmentation samples from the input image and generates an uncertainty map to quantify the degree of error in segmentation, so that erroneous representations are not utilized for T1 mapping. A morphological operation was then applied to detect the LV-RV intersection and delineate LV free wall from the interventricular septum. T1 ranges were obtained from the uncovered myocardial regions of interest with correction for T1 from the ventricular blood pools to improve discrimination between healthy subjects and those with cardiovascular disease (115). Using this proposed method, they characterized global and regional T1 values from over 10,000 subjects from the UKBB dataset which included a significant proportion of non-ischaemic cardiomyopathies. In line with present comprehension, they demonstrated that for those conditions in which diffuse fibrosis is more prevalent such as DCM, hypertrophic cardiomyopathy (HCM), and cardiac sarcoidosis, they found significantly higher T1 values (all p < 0.05). The quality control process is an important feature for clinical scalability of this tool and would enable this supplementary prognostic information to be added to each DCM case in a uniform manner with no added time-expense. Furthermore, it would enable large scale application to assess the role that native T1 analysis may have in deriving enhanced prediction of adverse outcomes for particular subgroups of DCM, particularly those who remain at risk despite having only mild or moderately impaired LV function. In order to be more generalizable, the proposed model requires validation on datasets acquired from the various different vendors available currently in clinical practice.