Assessing heterogeneity on cardiovascular magnetic resonance imaging: a novel approach to diagnosis and risk stratification in cardiac diseases

Abstract

Cardiac disease affects the heart non-uniformly. Examples include focal septal or apical hypertrophy with reduced strain in hypertrophic cardiomyopathy, replacement fibrosis with akinesia in an infarct-related coronary artery territory, and a pattern of scarring in dilated cardiomyopathy. The detail and versatility of cardiovascular magnetic resonance (CMR) imaging mean it contains a wealth of information imperceptible to the naked eye and not captured by standard global measures. CMR-derived heterogeneity biomarkers could facilitate early diagnosis, better risk stratification, and a more comprehensive prediction of treatment response. Small cohort and case–control studies demonstrate the feasibility of proof-of-concept structural and functional heterogeneity measures. Detailed radiomic analyses of different CMR sequences using open-source software delineate unique voxel patterns as hallmarks of histopathological changes. Meanwhile, measures of dispersion applied to emerging CMR strain sequences describe variable longitudinal, circumferential, and radial function across the myocardium. Two of the most promising heterogeneity measures are the mean absolute deviation of regional standard deviations on native T₁ and T₂ and the standard deviation of time to maximum regional radial wall motion, termed the tissue synchronization index in a 16-segment left ventricle model. Real-world limitations include the non-standardization of CMR imaging protocols across different centres and the testing of large numbers of radiomic features in small, inadequately powered patient samples. We, therefore, propose a three-step roadmap to benchmark novel heterogeneity biomarkers, including defining normal reference ranges, statistical modelling against diagnosis and outcomes in large epidemiological studies, and finally, comprehensive internal and external validations.

Graphical Abstract

Principle of novel cardiovascular magnetic resonance imaging–derived heterogeneity biomarkers to diagnose cardiac disease.

Open in new tabDownload slide

Introduction

Cardiovascular magnetic resonance (CMR) imaging is the reference standard for delineating cardiac structure, function, and in vivo tissue characterization. From a standard set of long- and short-axis balanced steady-state free precession (bSSFP) cine acquisitions, global measures of left ventricular (LV) end-diastolic volume (EDV), end-systolic volume (ESV), ventricular mass, stroke volume, and ejection fraction (EF) are readily calculated.^1,2 These biomarkers have normal ranges frequently defined as 2 standard deviations (SDs) above and below the population mean, adjusted for age, sex, and ethnicity; any measurement outside of this range may be considered pathological.³ A level of intra- and inter-individual heterogeneities is therefore expected and considered normal.

At the same time, the effects of different diseases across the heart are highly variable and heterogenous. Examples include focal septal or apical hypertrophy with reduced strain in hypertrophic cardiomyopathy (HCM) and replacement fibrosis, wall thinning, and akinesia in an infarct-related coronary artery territory. Myocarditis, dilated cardiomyopathy, and multi-system disorders including cardiac amyloidosis and sarcoidosis also show regional variability, characterized by more diffuse oedema, interstitial fibrosis, abnormal protein deposition, and granulomas, respectively. There is concern that conventional global cardiac biomarkers may be insensitive to the subtle changes in cardiac anatomy and physiology caused by especially early stages. Novel biomarkers that quantify this cardiac heterogeneity may aid the diagnosis of subclinical disease, more precisely risk stratification, and improve the prediction of treatment response (Figure 1).⁴

The aim of this article is to review proof-of-concept structural and functional cardiac heterogeneity biomarkers on CMR. We wish to first summarize emerging biomarkers, discuss their application in different clinical scenarios and cardiac diseases, and secondly, appraise current limitations and predict future directions.

Methods

A computerized search was performed through the PubMed database. The scope of the search was kept broad, and search terms were constructed around the themes of heterogeneity, biomarker, radiomics, strain, and CMR. Reflecting the novelty of the topic, many articles were identified through reference searches and discussions with experts in the field. The included articles were read in full; abstract-only articles were excluded. Key results were collated and finally discussed.

Overview of CMR

The detail, versatility, and reproducibility of CMR make it an ideal modality to assess cardiac heterogeneity. CMR provides unconstrained views of the entire heart and its relation to extra-cardiac structures, avoiding geometric assumptions. Exposed to varying radio signal sequences across different magnetic gradients in a typically 1.5 T magnetic field scanner, a patient’s heart re-emits signals that are processed to produce highly sensitive and specific cardiac images.¹ Cine and flow sequences delineate dynamic function with good spatial resolution.¹ Advanced sequences include pre- and post-vasodilator perfusion scans to investigate inducible ischaemia and early- and late-gadolinium enhancement (EGE and LGE) acquisitions to look for inflammation, fibrosis, and viability, respectively.¹ Parametric mapping techniques based on T₁ (native and contrast) and T₂ relaxation times are used to quantify changes in myocardial tissue composition (Table 1).¹ Importantly, previously seen as barriers to the widespread use of CMR, cost is falling and availability is increasing.¹

Strain and radiomics on cardiac imaging

Structural and functional heterogeneity assessment already exists in cardiac imaging. The most recognizable example is LV regional wall-motion abnormalities (RWMAs) on echocardiography.⁵ Strain is a relatively newer quantitative assessment of regional cardiac function that measures the degree of myocardial deformation along the longitudinal axis, the circumferential axis, and the radial short axis. Derived metrics include measures of dys-synchrony and discoordination (Table 2). On echocardiography, mechanical dispersion, defined as the SD of segmental time to peak negative longitudinal strain (LS) in a standard 16-segment LV model, is incrementally added to the risk of life-threatening ventricular arrhythmias in various cardiomyopathies, including hypertrophic, arrhythmogenic, ischaemic, and non-ischaemic forms.^13–17 A CMR study of 15 patients with non-ischaemic dilated cardiomyopathy also showed increased regional dispersion in LS and circumferential strain (CS) on tagged CMR.¹⁸

Radiomics is the archetype of heterogeneity assessment; it is based on the hypothesis that biomedical image data points reflect disease-specific changes.^8,9 Sophisticated digital algorithmic pattern recognition techniques model shape and tissue heterogeneity in a region of interest (ROI) from radiographic images.¹⁰ Examples are shape analysis to provide detailed morphometric information, histogram-based features to describe the distribution of voxel signal intensity (SI), and higher order texture analysis (TA) to objectively assess the pattern of voxel SI in the ROI (Table 2).^8–11 CMR images in Digital Imaging and Communications in Medicine or DICOM format are readily exportable onto open-source and commercially available radiomic analysis platforms. These include Py-Radiomics (Computational Imaging & Bioinformatics Lab, Harvard Medical School, Boston, MA, USA), Mazda (Institute of Electronics, Technical University of Lodz, Lodz, Poland), and TexRAD (Feedback Medical Ltd, London, UK). In a standard radiomic workflow routinely acquired clinical images undergo (i) segmentation; (ii) feature extraction; (iii) dimensionality reduction by principal component or cluster analysis; and (iv) logistic regression and more advanced machine-learning algorithms to build highly sensitive clinical models that are then internally and externally validated (Figure 2).^8–11 On computed tomography coronary angiography, unique radiomic signatures of coronary atherosclerotic plaque and perivascular inflammation improve risk stratification for future cardiovascular events.¹⁹

LV hypertrophy

Contemporary challenges

Undifferentiated LV hypertrophy (LVH) is a common clinical scenario with potentially life-altering implications for the patient and their family, depending on the underlying aetiology. Causes include physiological adaptation in athletes and remodelling secondary to intrinsic myocardial disease (e.g. HCM), systemic illnesses (e.g. amyloidosis) or increased pathological loading conditions (e.g. hypertension). Each condition is also very heterogeneous. In HCM, defined in probands by LVH ≥1.5 cm in ≥1 myocardial segments not explained by co-existent systemic disease, many genotype positive–phenotype negative patients live a normal lifespan without any symptoms. Conversely, 30–40% of HCM cohorts will experience an adverse event including sudden cardiac death (SCD).²⁰ Although sophisticated functional parameters (e.g. strain) and tissue characterization techniques (e.g. LGE) are very sensitive and specific to advanced stages of disease in the presence of significant regional hypertrophy and fibrosis, novel biomarkers for early diagnosis and phenotyping are critical to guide risk stratification and treatment.²⁰

Potential radiomic signatures

In a retrospective analysis involving 31 healthy controls and 185 individuals with LVH (50 HCM, 52 cardiac amyloid, 68 severe aortic stenosis (AS) and 15 hypertensive), histogram-based features on mid-short-axis unenhanced bSSFP cine slices identified radiomic signatures unique to the different pathologies.²¹ The variables included mean, SD, skewness, kurtosis, entropy, and mean positive pixel (MMP). Features from patients with HCM and cardiac amyloid demonstrated the greatest variability in healthy volunteers (P < 0.001; see Supplementary data online, Table S1).²¹ Their increased wall thickness is due to intrinsic myocardial disease, i.e. diffuse myocyte hypertrophy with focal fibrosis and myocardial disarray and increased extracellular volume secondary to abnormal protein deposition, respectively. LVH in AS and hypertension is a reaction to increased loading conditions and therefore has a less heterogeneous histopathology.

Application in hypertension

One important aim is to catalogue early LV remodelling secondary to chronic hypertension; tailored pharmacotherapy and treatment targets can then be offered to patients with and without the subclinical consequences of hypertension on imaging. In a UK Biobank retrospective case–control study of 200 cine images, including 100 with hypertension, an 11-variable radiomic model improved the discrimination between hypertensive and normotensive patients with an area under the receiver operating characteristic (AUROC) curve of 0.76 ± 0.13.¹⁰ Texture features were derived from different matrices, including the grey-level co-occurrence matrix (GLCM) (e.g. Homogeneity 1), grey-level run-length matrix (e.g. long-run emphasis), and grey-level size zone matrix (e.g. large area emphasis). In comparison, conventional imaging parameters, including LVEDV, ESV, and EF only, had an AUROC of 0.62 ± 0.09.¹⁰

Application in HCM

Pathognomonic patchy mid-wall LGE may be absent, sparse, or if identified, reflective of more advanced disease in HCM. The TA of short-axis non-contrast T₁-weighted slices from 32 HCM and 30 controls demonstrated significant differences in the features grey-level non-uniformity (GLNU), energy of wavelet coefficients in low-frequency subbands, fraction, and sum average irrespective of LGE (see Supplementary data online, Table S1).²² In a further case–control study of 12 HCM patients, GLNU and run-length non-uniformity (RLNU) on LGE sequences could discriminate non-hypertrophied non-fibrotic segments from healthy controls (see Supplementary data online, Table S1).²³ HCM also leads to reduced global LS, CS, and radial strain.^24,25 To measure heterogeneity in function across the myocardium of 22 HCM patients, Sakamoto et al. derived the coefficients of variation (CoV) of regional LS and CS from a 16-segment LV model (see Supplementary data online, Table S2). LS_CoV was significantly greater in HCM vs. healthy volunteers and CS_CoV had 83% sensitivity and 94% specificity to identify extensive LGE (%LGE ≥15%), a marker of increased SCD risk in HCM.²⁵ The diagnostic sensitivity of CMR for subclinical HCM may improve if radiomics and strain can reliably measure early cardiomyocyte disarray and dysfunction.

Myocarditis: a heterogeneous diagnosis

Contemporary challenges

Myocarditis is notoriously difficult to diagnose. The reasons for this include the varied distribution of the affected myocardium (e.g. focal vs. diffuse) and the heterogenous clinical course (e.g. progressively worsening shortness of breath vs. fulminant cardiogenic shock). Diagnosis is also time-dependent, as pathognomonic changes on imaging (e.g. oedema) and histopathology (e.g. inflammatory cell infiltrates) regress with time as myocarditis self-resolves or progresses to a dilated cardiomyopathy.²⁶ Endomyocardial biopsy (EMB) is the diagnostic gold standard, but the invasive procedure carries significant risks (e.g. ventricular perforation).^27,28 It also has an unacceptably high false-negative rate due to the patchy nature of the disease, even when using Dallas criteria with immunohistochemical analysis and polymerase chain reaction test for microbial detection.^29,30 On CMR, findings consistent with acute myocarditis are at least two of three Lake Louise Criteria (LLC): (i) raised focal or global T₂-weighted SI, (ii) increased global EGE, and (iii) subepicardial fibrosis on LGE.²⁶ Updated LLC now include T₁ and T₂ parametric mapping techniques; however, their high intra- and inter-individual variabilities make discrimination between health and disease challenging.^26,31

Potential radiomic signatures

In a 79-patient substudy of the Magnetic Resonance Imaging in Myocarditis (MyoRacer) trial, the diagnostic utility of TA was compared with LLC and global native T₁ and T₂ in differentiating EMB-positive from EMB-negative acute and chronic heart failure (HF)-like myocarditis.³² Of the conventional CMR parameters, T₂ had the highest AUROC curve in both forms of presentation (i.e. 0.69 and 0.62, respectively).³² After dimension reduction and feature selection, the most predictive texture features were T₂-kurtosis (AUROC 0.81) and T₁-GLNU (AUROC 0.74) in chronic myocarditis and T₂-GLNU (AUROC 0.69) in acute myocarditis (see Supplementary data online, Table S1).³² Combinations of T₂-kurtosis and T₁-GLNU (AUROC 0.85) in chronic myocarditis and of global native T₂ and T₂-GLNU in acute myocarditis (AUROC 0.76) had the highest combination of sensitivity and specificity for EMB-positive myocarditis.³² A similar analysis was conducted on a further 44 patients from the MyoRacer trial, presenting with infarct-like acute myocarditis.³³ The combination of T₂-RLNU and GLNU was better than LLC and global native T₁ and T₂ values in discriminating EMB-positive from EMB-negative infarct-like acute myocarditis (AUROC 0.88 vs. max 0.65; see Supplementary data online, Table S1).³³

CMR definition of myocardial infarction

Contemporary challenges

CMR with LGE can accurately quantify the extent of myocardial infarction (MI), but gadolinium-free imaging techniques are desirable due to frequent co-existent chronic kidney disease and the risk of nephrogenic systemic sclerosis.^12,34,35 LGE sequences are also limited in their ability to differentiate acute from chronic MI and use an oversimplified dichotomization of <50% transmural LGE to define myocardial viability, two important considerations in the treatment and revascularization of multi-vessel coronary artery disease.¹²

Potential radiomic signatures

The TA of cine CMR sequences from 120 patients with LGE-proven MI and 60 controls demonstrated significant differences in 2 histogram-based features [first percentile (perc0.01) and variance], one second-order texture feature {sum entropy [S(5,5)SumEntrp]}, and two higher-order features [teta1 and energy of wavelet coefficients in high-frequency subbands (WavEnHH.s-3)] (P < 0.001; see Supplementary data online, Table S1).³⁴ The AUROC for the two most discriminative features, perc.01 and teta1, combined was 0.92.³⁴

After feature extraction and dimension reduction, Larroza et al.³⁵ found that 9 of the 10 best discriminators between acute and chronic MIs on cine sequences were derived from the GLCM with an average AUROC >0.7 in most test sets. In a similar analysis, the same group of authors concluded that local binary pattern features on cine images could discriminate between likely viable (<50% transmural LGE) and likely non-viable (≥50% transmural LGE) myocardial segments with an AUROC >0.8.¹²

Radiomic and strain biomarker limitations

Future structural radiomic and functional strain-based CMR biomarkers need (i) to have a clearly defined reproducible derivation method, (ii) to add new incremental information about the disease and its pathogenesis, and most importantly, (iii) to facilitate improved management from diagnosis through risk stratification to treatment response.^36,37

Multiple parameters can affect CMR acquisitions, including magnetic field strength, spatial resolution, signal-to-noise ratio, slice thickness and orientation, ROI demarcation, and the selected sequence with or without contrast. By directly altering voxel size, strength and pattern imaging protocols can profoundly impact on radiomic, i.e. textural features.^8,21,32,33 Another caveat specific to radiomics is the high number of features analysed in small patient cohorts without internal or external validations, increasing the risk of Type 1 statistical errors or false positives.^9,21,32,34 Although readily computable through open-source software, a higher-order texture feature may be less intuitive and difficult to rationalize, thus hindering its widespread adoption.²¹

In comparison with EF and RWMA, strain-derived metrics more objectively quantify regional and global cardiac function. However, values and reference ranges remain dependent on the CMR strain technique [e.g. myocardial tagging, displacement encoding with stimulated echoes (DENSE) and feature tracking], post-processing software, operator experience, and image quality.^13–17 There are also indirect effects of cardiac pathology and its treatment through blood pressure and heart-rate changes on strain measurements.^13–17

As simple as the mean absolute deviation

A simple, yet promising heterogeneity biomarker is the mean absolute deviation of log_e-transformed segmental pixel SDs (madSDs) from their overall average in a standard 16-segment LV model on native T₁ and T₂ mapping sequences. The central dogma of heterogeneity biomarkers is that cardiac disease can be focal, patchy, diffuse, or a combination thereof. By averaging across the entire myocardium, global measures of remodelling (e.g. LV mass and EDV), function (e.g. EF), and tissue characterization (e.g. global native T₁ and T₂) are blunted to the early small changes, especially the pre-morbid state. Instead, madSD amplifies tissue heterogeneity by first measuring pixel dispersion at a segmental level with SD and then at a global level with mean absolute deviation.

Two retrospective case–control studies of 31 and 68 patients with CMR and clinical diagnoses of acute myocarditis, respectively, compared the diagnostic power of T₂-derived madSD and similar metrics to more conventional parameters.^31,38 First, global myocardial T₂ relaxation times and their SD were greater in patients than healthy volunteers, but still within the normal reference ranges.^31,38 Secondly, mean absolute deviation of segmental T₂ (madT2) and SD (madSD) from their overall myocardial mean on a 16-segment LV model were better able to discriminate between acute myocarditis and healthy volunteers (see Supplementary data online, Table S1 and Figure 3).^31,38 The combination of madSD and maxT2 in the combined cohort of 99 patients was as powerful as LLC in diagnosing acute myocarditis with 75% sensitivity and 80% specificity.^31,38 Perhaps the single greatest argument for madSD as the next heterogeneity biomarker is that it is a ‘relative’ parameter, measuring an individual’s deviation from their own mean. Inherently, it is less affected by high intra- and inter-individual variabilities of T₂ times and bias introduced by uncontrolled factors (e.g. different CMR imaging protocols and patient compliance). Other important advantages are non-reliance on intravenous contrast administration and subjective image interpretation.

In one of the first studies to link cardiac heterogeneity biomarkers to hard clinical endpoints, Nakamori et al.³⁹ derived madSD from native T₁ parametric mapping sequences in a prospective cohort of 115 non-ischaemic cardiomyopathy patients referred for primary prevention implantable cardioverter defibrillators (ICDs; see Supplementary data online, Table S1 and Figure 3). A combination of native T₁ > 2 SD above the mean of healthy controls and T₁ madSD >0.24 (median) was at least comparable with the presence, location, and extent of LGE in predicting arrhythmic events.³⁹ The underlying hypothesis is that increased tissue heterogeneity secondary to interspersion of necrotic and viable myocardium with variable conduction velocities facilitates re-entry circuits, increasing the likelihood of malignant arrhythmias. This also explains why the extent of peri-infarct or grey zone on LGE, defined by SI 2–3 SD above the peak SI of the remote myocardium or greater than the peak remote SI, but <50% of the peak SI in the infarct core, is more predictive of appropriate ICD therapy and/or ventricular arrhythmias than the extent of the infarct core.^40,41 The hope is that combination of T₁-derived madSD and established risk modifiers, including severe LV systolic dysfunction (LVSD; LVEF ≤35%), genotypes, and clinical phenotypes can substantially reduce the 25% inappropriate shock burden in patients with primary prevention ICDs.^39–42

Measuring dys-synchrony through time

Ventricular electromechanical dys-synchrony predicts cardiac decompensation and SCD in heart disease.^6,7,43,44 Although a crude measure of this functional heterogeneity or dys-synchrony with poor prognostic power, QRS prolongation (≥120 ms) guides cardiac resynchronization therapy (CRT) in HF.^43,44 Alternative biomarkers are strain-derived dys-synchrony parameters, including SD of time to peak negative strain in all myocardial segments (TTP_SD). These time-based metrics may be more reproducible and less susceptible to the differential effects of technical, operator, and patient factors than basic strain parameters.^6,7

Similar to TTP_SD CMR-tissue synchronization index (CMR-TSI) measures contractile heterogeneity by quantifying the SD of time to maximum segmental radial wall motion in a 6 × 8 segmented LV short-axis stack model (see Supplementary data online, Table S2 and Figure 4).⁴⁴ In a prospective study of 77 patients with severe LVSD, CMR-TSI ≥110 ms at least doubled the risk of death and hospitalization for HF post-CRT implantation. Versus patients with CMR-TSI <110 ms, this subgroup did not have LV remodelling.⁴⁴ CMR-TSI is an example experimental functional heterogeneity biomarker that can predict risk and response to treatment.

Heterogeneity biomarker roadmap

In summary, two contending structural and functional CMR heterogeneity biomarkers are log_e-transformed madSD and CMR-TSI based on a segmented LV model, respectively. We propose a three-step roadmap to comprehensively benchmark the biomarkers.

1. Biomarker definition: Normal reference ranges, defined by 2 SD above and below the 95% confidence intervals of the population mean need to be derived from CMR imaging of healthy volunteers.³ Paired studies can assess intra- and inter-observer variabilities as well as test–retest reproducibility. Other important considerations include the choice of CMR sequences (e.g. cine, T₁, or T₂ mapping); scanner vendors, field strength, and operator expertise; and the cardiac chamber, ROI, and number of segments to be included.

2. Role in clinical assessment and management: Epidemiology studies (e.g. UK Biobank and Jackson Heart Study) with comprehensive imaging, including CMR, are now accessible to researchers.⁴⁵ These population cohorts are well-phenotyped at baseline and continue to collect extensive outcome data.⁴⁵ Applied simple test statistics, regression, and survival models can evaluate the role of the biomarkers in diagnosis and risk stratification.

3. Internal and external validations: The heterogeneity biomarkers should be vigorously tested internally and externally in said population studies that have recruited thousands of volunteers with CMR data as well as patient cohorts.⁴⁵ Finally, their clinical utility should be assessed in a randomized controlled trial setting.

Conclusion

Disease increases heterogeneity in cardiac tissue composition and physiology; this is seen in ischaemic heart disease, HCM, myocarditis, and dilated cardiomyopathy. The advancing technology of CMR, including strain-based assessment and post-processing radiomics, makes it an ideal modality to assess cardiac heterogeneity at a structural and functional level. Small cohort and case–control studies have investigated proof-of-concept CMR heterogeneity imaging biomarkers that may facilitate early diagnosis and improved risk stratification. Two of the most promising are madSD and CMR-TSI, which use voxel dispersion and contraction dys-synchrony as markers of histopathology and reduced heart function, respectively. We describe a three-step benchmarking plan to define, calibrate, and validate these and other novel heterogeneity biomarkers.

Author contributions

K.H.: methodology (lead), investigation (lead), data curation (lead), formal analysis (lead), visualization (equal), writing—original draft preparation (lead), and writing—review and editing (equal). M.Y.K., C.A.A.C., and N.A.: supervision (equal) and writing—review and editing (equal). G.S.D.: visualization (equal). S.E.P.: conceptualization (lead), supervision (equal), and writing—review and editing (equal).

Supplementary data

Supplementary data are available at European Heart Journal - Cardiovascular Imaging online.

Funding

K.H. was supported by a British Heart Foundation Clinical Research Training Fellowship (FS/CRTF/23/24428).

Data availability

No new data were generated or analysed in support of this research.

References

Author notes