Towards accurate and precise T1 and extracellular volume mapping in the myocardium: a guide to current pitfalls and their solutions

Mapping of the longitudinal relaxation time (T 1) and extracellular volume (ECV) offers a means of identifying pathological changes in myocardial tissue, including diffuse changes that may be invisible to existing T 1-weighted methods. This technique has recently shown strong clinical utility for pathologies such as Anderson-Fabry disease and amyloidosis and has generated clinical interest as a possible means of detecting small changes in diffuse fibrosis; however, scatter in T 1 and ECV estimates offers challenges for detecting these changes, and bias limits comparisons between sites and vendors. There are several technical and physiological pitfalls that influence the accuracy (bias) and precision (repeatability) of T 1 and ECV mapping methods. The goal of this review is to describe the most significant of these, and detail current solutions, in order to aid scientists and clinicians to maximise the utility of T 1 mapping in their clinical or research setting. A detailed summary of technical and physiological factors, issues relating to contrast agents, and specific disease-related issues is provided, along with some considerations on the future directions of the field.


Introduction
Mapping of the longitudinal relaxation time, T 1 , and extracellular volume (ECV) in the human heart has recently shot to prominence on the merits of the modified Look-Locker inversion recovery (MOLLI) imaging sequence and related techniques [1]. These methods allow quantitative tissue characterisation in the myocardium, adding new information to that provided by T 1 -weighted techniques such as late gadolinium enhancement (LGE) imaging. For focal fibrosis, LGE provides excellent delineation of lesions with some means of quantifying their volume; however, LGE does not give a T 1 estimate and may not be able to identify widely distributed or diffuse myocardial diseases. For example, when LGE is applied in diffuse fibrosis, the myocardium can appear isointense and indistinguishable from normal myocardium, as multiple uncalibrated factors affect the image brightness. These are clear limitations of LGE techniques, and in such situations quantitative T 1 and ECV mapping is advocated [2,3]. Use of myocardial T 1 mapping is now widespread, with most MRI manufacturers offering T 1 mapping solutions. However, great care must be taken when applying these methods clinically, given the need for protocol optimisation and locally-derived normal ranges.
The accuracy and precision of myocardial T 1 mapping has been the focus of several studies to date [4][5][6][7], and has been discussed to some degree in much of the literature. Innovations in the field are considered in terms of their impact on accuracy and precision, typically offering trade-offs in one or the other for faster or more-accommodating scans. However, there are still several longstanding pitfalls associated with T 1 and ECV mapping that affect accuracy and precision, and this review aims to give a comprehensive description of these, along with potential solutions, with the intent of aiding physicists 1 3 and clinicians to maximise the clinical utility of T 1 mapping. Indications will also be given as to what is reasonably achievable with myocardial T 1 mapping in specific clinical applications, without a full clinical review, for which the reader is directed to Haaf et al. [8], Taylor et al. [9], Kammerlander et al. [10], and Schelbert and Messroghli [11], amongst others. The fundamentals of T 1 mapping methods, the available pulse sequences, and the history of the technique will be discussed briefly, but again readers are directed to more-detailed reviews for full technical information: for example, by Kellman et al. [5] and Higgins and Moon [12]. Finally, the consensus statement of the Society for Cardiovascular Magnetic Resonance (SCMR) and the CMR Working Group of the European Society of Cardiology can be consulted for recommendations on how to set up a robust T 1 mapping protocol [13]. A follow-up parametric mapping consensus statement from the SCMR and the European Association for Cardiovascular Imaging (EACVI) is in preparation.
This review will focus on sources of bias and variability in myocardial T 1 and ECV estimation: first, technical and physiological pitfalls; second, issues relating to contrast agents; and third, specific disease-related issues. It will conclude with future directions of the field and some summary recommendations, including guidance on how T 1 estimation accuracy might be traded for increased clinical utility. Within each section, pitfalls will be introduced, their mechanisms described, and suggestions offered for how to mitigate or eliminate them, along with possible future solutions, where available. For ease of reference, Table 1 lists the technical and physiological pitfalls discussed in this review in the order they are introduced in the text, with summaries of their relative effects on T 1 and ECV mapping accuracy and precision.

A brief introduction to T 1 mapping methodology
All routinely available T 1 mapping methods currently rely on preparing the longitudinal magnetisation using inversion or saturation radiofrequency (RF) pulses, applied to the whole imaging volume. Many pulse sequences for T 1 mapping can be grouped according to the magnetisation preparation: inversion recovery sequences, including MOLLI [1,14] and shortened MOLLI (ShMOLLI) [15]; saturation recovery sequences, including independent saturation recovery single-shot acquisitions (SASHA) [16], saturation method using adaptive recovery times for cardiac T 1 mapping (SMART 1 Map) [17], and short acquisition period T 1 (SAP-T 1 ) [18]; and mixed preparation sequences, such as saturation pulse-prepared heart-rate independent inversionrecovery (SAPPHIRE) [19]. In practice, a mixture of magnetisation preparations and T 1 -weighted image acquisitions are performed during a breath-hold, over several cardiac cycles. The aim is to sample the T 1 recovery over a range of delay times, and pixel-by-pixel curve-fitting is used to estimate T 1 values. This produces a pixelwise T 1 map, often with other output maps as quality indicators; see Kellman et al. for a flowchart illustrating the pipeline of T 1 and ECV map generation [20]. Most MRI manufacturers provide inline software for map calculation, but open source tools are also available [21,22].
Non-mapping approaches can also estimate T 1 in the heart: the inversion-recovery cine sequence, also known as the Look-Locker cine (LL-cine) technique [23,24], relies on averaged signal intensities over regions of interest (ROIs) for T 1 curve-fitting. This approach uses spoiled gradient recalled echo (GRE) cine, avoiding the factors affecting single-shot balanced steady-state free-precession (bSSFP). Note that many studies that have used the LLcine method have repeated it at multiple washout times, as this improves ECV accuracy [24].
Given its popularity, the original MOLLI sequence will be considered the default method in this review, with other T 1 mapping methods and schemes being addressed where appropriate. A common nomenclature for T 1 mapping schemes will also be adopted [5]. This notation lists the number of images acquired following a magnetisation preparation pulse, along with the free-recovery, inter-inversion pause in brackets. Timings are given in beats, "b", or seconds, "s". For example, 3b(3b)3b(3b)5b uses three Look-Locker sets, of three beats, three beats, and five beats, respectively, with pauses of three R-R intervals between each set. A 5s(3s)3s scheme uses two Look-Locker sets and a minimum pause of 3 s between these, rounded up to the next whole R-R interval.

Look-Locker correction
The original 3b(3b)3b(3b)5b MOLLI sequence [1] and subsequent optimised versions rely on Look and Locker's correction for rapid T 1 estimation: namely, for sampling of the recovering longitudinal magnetisation with a series of small-flip-angle excitation pulses [25,26]. The magnetisation is perturbed by these excitation pulses, causing flattening of the recovery curve, and yielding an apparent T 1 , known as T 1 *, when curve fitting: where S(t) is the signal at time TI after inversion. The "true" T 1 is usually longer than T 1 * and can be calculated using the so-called Look-Locker correction:

3
The Look-Locker correction assumes continual repetition of a small flip angle spoiled GRE readout. MOLLIbased techniques violate the Look-Locker assumptions by using: (1) a bSSFP readout, which is sensitive to T 2 and, weakly, to magnetisation transfer (MT), unlike spoiled GRE; (2) a relatively large excitation flip angle, 35° at 1.5T; and (3) an intermittent sampling scheme, governed by heart rate for original MOLLI, to maximise the spread of inversion recovery times. Factors (1) and (2) lead to progressive saturation of the longitudinal magnetisation, causing a negative T 1 bias even after Look-Locker correction. Furthermore, Look-Locker sets are separated by pause intervals, aiming to allow sufficient recovery of the magnetisation prior to the next inversion pulse. If any of these pause intervals are too short for the application, be it native or post-contrast, this can lead to additional bias in T 1 estimates. This can be particularly problematic for fast heart rates, as well as longer T 1 values, such as in native myocardium at 3T.
Bias resulting from limited magnetisation recovery can be mitigated by using an optimised MOLLI acquisition scheme: by extending the inter-inversion pause in beats or by specifying pauses in seconds rather than beats [4]. The 2013 SCMR consensus recommends a 5s(3s)3s scheme for native T 1 mapping and a 4s(1s)2s(1s)1s scheme for postcontrast acquisitions [13].
Alternatively, saturation recovery methods such as SASHA [16] sample the recovering longitudinal magnetisation using an independent preparation in each heartbeat, obviating the need for Look-Locker correction over multiple shots at the expense of a loss in precision due to a smaller dynamic range of T 1 recovery. The saturation recovery is still affected during each SASHA single-shot readout, but this produces negligible bias in T 1 estimates obtained from curve fitting of the independent saturationrecovery images [16,27].

Partial volume
Partial volume of different tissues within a voxel is prevalent in T 1 mapping with single-shot imaging, where inplane spatial resolution is necessarily somewhat coarse. Furthermore, the endo-and epi-cardial borders of the myocardium are often oblique to the imaged slice, especially for planes far from the mid-ventricle or in abnormal ventricles. Partial volume can lead to substantial errors, which motivates careful ROI delineation on T 1 maps [28].
Partial volume causes bias in the apparent T 1 , especially when the tissues included in a voxel have strongly different T 1 values. Prominent effects occur at the endocardial border where the difference in T 1 between blood and myocardium leads to overestimation of native T 1 estimation in subendocardial voxels. Partial volume also occurs between myocardium and other tissues, most often fat, whose chemical shift causes variable bias in the pixel T 1 [29]; this has clinical relevance, and is discussed further in the "Errors in specific clinical applications" section.
Finer in-plane resolution of the T 1 mapping sequence would theoretically reduce partial volume effects, but demands longer single-shot imaging duration if no tradeoffs are made, increasing the risk of cardiac-motion blurring. Parallel imaging and partial Fourier in the phaseencode direction are commonly employed to allay this problem. It is essential to avoid partial volume when drawing regions of interest (ROIs), which are often limited to mid-wall when possible [28,30]. Cardiac motion during the single-shot imaging also corrupts myocardial signal with blood signal in less obvious ways, and is discussed later.
"Black-blood" T 1 mapping aims to eliminate blood partial volume for improved T 1 estimation accuracy [31], whereby the magnetisation of flowing blood is nulled by motion-sensitive dephasing immediately before each single-shot image. Multiecho fat-water-separated imaging strives to separate fat signal from the thin RV [32], and has been combined with the same method of blood suppression [33]. However, these approaches are not routinely reliable.
To achieve finer spatial resolution, and thus reduce the impact of partial volume in source images on pixelwise mapping, segmented k-space image acquisition over multiple cycles is required, and is often combined with some undersampling strategy [34,35]. Segmented acquisition is severely affected by R-R variability in inversion-recovery methods, and is very slow to acquire the fully recovered image in saturation-recovery methods. The non-mapping LL-cine approach can acquire fine spatial and temporal resolutions, but cardiac motion occurs during recovery, raising questions about the impact of through-slice motion, and substantial post-processing labour is required to optimise ROIs used for curve-fitting and regional estimation of T 1 .

Factors affecting magnetisation preparation pulses
The fitting models used in T 1 mapping usually assume exact inversion or saturation of the longitudinal magnetisation, or fit an extra parameter instead, reducing precision. Another approach uses prior knowledge of an inversion factor, which can be estimated from fully relaxed reference images to correct the estimated T 1 [40][41][42].
Conventional RF pulses require accurate RF transmit (B 1 ) fields to achieve their prescribed flip-angle, so are sensitive to B 1 inhomogeneity, which can be substantial across the heart, particularly at 3T [40,43]. The flip-angle achieved by conventional non-selective RF pulses is also affected by off-resonance errors due to B 0 inhomogeneity. Therefore, adiabatic inversion pulses or composite saturation pulses are widely used for mapping to reduce sensitivity to both B 0 and B 1 inhomogeneity [40,44]. However, the longer duration of some adiabatic pulses can increase T 2 decay during their execution and introduce sensitivity to off-resonance phase accumulation. Kellman, Herzka, and Hansen suggest the use of a relatively short tan/tanh adiabatic pulse for optimal inversion efficiency [40] (Fig. 1), while a composite saturation design is recommended for saturation-recovery methods such as SASHA, provided the higher specific absorption rate and pulse duration are acceptable [44]. Optimised saturation precision, or efficiency, is important in SASHA for two reasons: (1) to enable two-parameter curve-fitting for T 1 estimation, as opposed to fitting of saturation efficiency as a third parameter; and (2), reliable removal of any history effect from previous cycles by nulling the longitudinal magnetisation prior to each independent shot.
Although the aforementioned RF pulse designs are more tolerant to B 0 and B 1 errors, for mapping it is vital to optimise both B 0 and B 1 over the relevant volume. See Fig. 2 for examples of B 0 and B 1 maps, in vivo. The advent of RF (B 1 ) shimming hardware with optimised volume calibration methods has enabled substantial improvements in B 1 uniformity over the heart [45] even with only two whole-body transmitter channels [46].

Inversion-recovery multishot bSSFP
Most T 1 mapping methods use bSSFP to sample the recovering longitudinal magnetisation, as this method offers a higher signal-to-noise ratio (SNR) than spoiled GRE. However, bSSFP is sensitive to T 2 and off-resonance, where both sensitivities are modulated by excitation flipangle, RF pulse repetition time (TR), and magnetisation transfer [47]. These factors have greater impact on estimated T 1 when the bSSFP flip angle is higher or the TR is longer, and native myocardial T 1 values are more affected than the shorter post-contrast T 1 values. Off-resonance is larger at 3T, mitigated by use of a lower flip angle [48], but reducing T 1 error through use of a shorter TR is largely limited by patient peripheral nerve stimulation [49].
Another consideration with bSSFP is the transient period before the steady-state is established, which is characterised by oscillatory magnetisation, causing image artefacts. The intensity of the oscillations depends on both the bSSFP catalysation used to stabilise the signal, and the k-space trajectory [50].

Catalysation sequences for bSSFP
For T 1 mapping, some single-shot bSSFP images at short inversion-recovery times are required for optimal curve-fitting, although for SASHA a later start is preferred, as earlier readouts have low SNR [27]. Short delay times prevent stabilisation of the bSSFP signal before phase-encoded data acquisition commences, and since all shots should be acquired with identical parameters, a longer stabilisation for the later inversion-recovery shots is inadvisable. The "design region", enclosing the maximal likely in vivo cardiac B 0 and B 1 distortion, is represented by the dotted green box. Adapted with permission from Kellman et al. [40] Several catalysation or "priming" sequences are used in bSSFP to accelerate the signal's approach to the steady state, and these are particularly important in T 1 mapping's limited time window before image data must be acquired, especially for centre-out k-space ordering, which is discussed later. Catalysation details may be concealed from the scanner's user interface, and may change through software upgrades without warning. Schemes include halfalpha [51], and linear ramp-up [52] (see Fig. 3); the type and duration of the scheme affects T 1 estimation accuracy and precision, with bias errors of the order of 5% or more for some approaches [53].
The linear sweep up scheme is a good approach, as this sufficiently quells oscillatory behaviour prior to k-space filling [53].

Impact of T 2 on estimation of T 1 by bSSFP readouts
Shorter T 2 relaxation times are associated with underestimation of T 1 due to the T 2 /T 1 weighting of bSSFP [5,43], and their effect on preparation pulses. The Look-Locker sets of MOLLI sequences are more sensitive to T 2 than the independent images used in saturation methods like SASHA, due to accumulated T 2 -related saturation between single-shot images in each set [54]. When T 2 is long, such as in left ventricular blood, MOLLI estimation of T 1 increases towards the true value.
In addition to low flip angles and short TR, coarser acquired resolution in tandem with parallel imaging and partial Fourier in the phase-encode direction can reduce the length of the bSSFP pulse train and thus mitigate T 1 estimation bias resulting from T 2 -related saturation.

Impact of off-resonance on bSSFP readouts
A disadvantage of bSSFP is its sensitivity to off-resonance [55], which causes dark banding artefacts and associated T 1 errors; however, off-resonance errors in estimated T 1 can also occur in myocardial regions without banding artefacts. The inferolateral segment is particularly vulnerable to this problem. Furthermore, the off-resonance sensitivity of bSSFP is also influenced by the catalysation sequence used, as shown in Fig. 3.
The sensitivity to off-resonance is not identical across all of the bSSFP source images. Given that single-shot bSSFP is not fully stabilised for T 1 mapping sequences, the impact of off-resonance varies at different points on the T 1 recovery curve, and so does not cancel out of the curve-fitting estimation of T 1 [48]. ECV measurements are less strongly affected by off-resonance than native T 1 estimates, showing a bias error of around 1% or less [48].
Off-resonance errors can be reduced through volume B 0 shimming over the heart and great vessels; however, even if second-order B 0 shimming is carefully optimised, it cannot correct very local B 0 distortions (Fig. 2a). Investigators should familiarise themselves with B 0 shimming routines on their scanner, considering factors such as cardiac gating and respiration, among others. Projection-based shimming algorithms [56] are widely available; however, image-based shimming methods [57,58] may offer better control over local B 0 , and they provide B 0 fieldmaps that can assist quality control of T 1 studies.
Off-resonance distortion of T 1 can be reduced by lower flip-angles and/or a shorter TR. For example, a coarser frequency-encode resolution may reduce TR; however, this shimming and dual-channel RF calibration. Note the distinct inhomogeneity of B 0 near the coronary veins in a (arrows), and slightly reduced pulse performance across the right ventricle in b, where the measured B 1 drops to around 50-60% of the nominal value may also automatically modify the phase-encode resolution, reducing the number of RF pulses before the centre of k-space and affecting stabilisation. Furthermore, coarser frequencyencode spatial resolution increases partial volume. Any such changes require attention with regards to effects on estimated T 1 and possible invalidation of normal range data [13].
A recent development substituted MOLLI's bSSFP readout with a spoiled GRE sequence [59], which avoids the more complex sensitivities of bSSFP, improves T 1 estimation accuracy, and reduces the sensitivity of T 1 estimation to T 2 . It has also been applied in patients with implanted devices that cause severe off-resonance artefacts, precluding bSSFP imaging [60]. However, these benefits come with several disadvantages: spoiled GRE shows reduced SNR versus bSSFP; its T 1 estimation precision is also reduced; and adequate spoiling may be difficult. Spoiling has been shown to be problematic for the variable flip angle method [61], but may be less so for inversion-recovery based T 1 mapping [59].

Schemes for k-space filling
To date, most T 1 mapping has used linear phase-encode ordering, where phase-encoding gradient amplitudes are stepped through incrementally. This avoids eddy-currentrelated signal perturbations, but causes progressive T 2related saturation in the approach to the centre of k-space, leading to T 1 underestimation. Alternatively, centric phaseencode ordering, also known as the centre-out or low-high approach, fills k-space from the centre outwards with alternating and increasing phase-encoding gradient amplitudes; this avoids the T 2 -related saturation of linear ordering at the expense of increased eddy-current-related artefacts [53].
Although linear-ordering is typically used, several other phase-encode ordering schemes have been investigated to date [43,53,62]. Paired phase-encoding has been proposed to mitigate the artefacts associated with centric ordering [63]; however, it has shown mixed results for T 1 mapping [53,62], performing well only with longer catalysation schemes.

Signal-to-noise
All T 1 mapping methods acquire multiple T 1 -weighted source images, each of which has its own SNR per tissue, and the noise level will influence the sampled points during curve-fitting. The fewer T 1 -weighted source images used to reconstruct a T 1 map, the poorer the curve-fit conditioning Fig. 3 The effect of balanced steady-state free-precession catalysation sequences on native T 1 maps. Plots show the effect of linear ramp-up (a) and half-alpha (b) catalysations on the magnetisation evolution (i) and frequency response at the centre of k-space (ii) of balanced steady-state free-precession at 3T. Short axis, mid-ventricu-lar native T 1 maps are shown for each method (iii). Simulations were performed with a repetition time of 2.8 ms, a flip angle of 35°, a T 1 /T 2 of 1200/45 ms, and 10 catalysation pulses for each method, with a further 39 pulses before the centre of k-space and the poorer the T 1 estimation precision [15]. Given the limited number of shots taken throughout longitudinal recovery, their optimum distribution in comparison to the relevant range of T 1 is also important, motivating different sequence schemes for native and for post-contrast mapping [5,27].
SNR varies spatially across the heart, predominantly due to the sensitivity profile of the receiver coil array. Low SNR is most evident in the lateral wall, which is farthest from the coil, and thus this region is more prone to noise-related T 1 estimation bias and dispersion [4]. This effect is in addition to susceptibility artefact seen in the lateral wall-another reason why clinical T 1 measurements for assessment of diffuse fibrosis are often confined to the interventricular septum [30].
Imaging at higher field strengths can mitigate errors resulting from low SNR, as shown by Piechnik et al. [15], who reported approximately 30% reduction in coefficients of variation for MOLLI and ShMOLLI T 1 estimates when moving from 1.5 to 3T. Conversely, 3T exacerbates off-resonance and B 1 inhomogeneity effects, though their impact can be controlled.

Influence of field strength
In addition to the aforementioned off-resonance and B 1 inhomogeneity issues at higher field strengths, and the potentially increased SNR, there are also differences in native T 1 and T 2 values between 1.5 and 3T.
A large multicentre study of native T 1 and ECV values in normal myocardium, using original 3b(3b)3b(3b)5b MOLLI, reported mean (standard deviation) native T 1 values of 950 (21) ms at 1.5T and 1052 (23) ms at 3T, and mean (standard deviation) ECVs of 0.25 (0.04) at 1.5T and 0.26 (0.04) at 3T [64]. The increased T 1 at 3T can lead to insufficient longitudinal recovery between Look-Locker sets, causing T 1 underestimation; furthermore, reduced myocardial T 2 at 3T relative to 1.5T introduces additional negative bias due to the T 2 /T 1 weighting of bSSFP and signal decay during preparation pulses [5].
Increased B 0 and B 1 inhomogeneity at 3T can be mitigated using appropriate B 0 and RF shimming, respectively. Shallower excitation flip angles can also allay these effects, if the increased SNR at 3T is traded off [5].

Breath-holding
Currently, T 1 mapping requires breath-holding to minimise respiratory motion while source images are acquired. Original MOLLI used a breath-hold duration of approximately 17 cardiac cycles [1], while newer variants require around 10 or 11 s [4]. A shorter breath-hold is an advantage in routine work [15], but it causes a reduction in SNR. Imperfect breath-holds typically lead to misregistered source images, which corrupt the set of signal intensities used for pixelby-pixel curve fitting and, in turn, decrease T 1 estimation accuracy and precision. Often, misregistration is not readily apparent in calculated T 1 maps, unless a confidence-map is provided alongside or overlaid on the T 1 map. See Fig. 4 for an example of this.
Post-acquisition quality control has some aspects in common for respiratory motion and cardiac misgating or arrhythmia: T 1 mapping source images should be examined carefully for displacements, even if motion-corrected images are also available. Mislocated tissue in or through the selected slice, for any of the shots, may preclude correct mapping of localised disease, such as myocarditis. If significant displacements are found, nonrigid registration may register most myocardial pixels [20,65], but cannot correct through-slice displacement. Input image registration in T 1 mapping is challenging due to the large image-contrast variations between source images, and can be unreliable when tissues are imaged near the null point of longitudinal magnetisation recovery. While there are strategies for dealing Fig. 4 An example of motion-related T 1 estimation error shown on quality control maps. A short-axis native T 1 map corrupted by respiratory motion (a) demonstrates excessive pixelwise curve-fitting residual errors, indicated on a confidence map (b) by the addition of marked pixels to the same map shown in a with this issue [65], motion-corrected images should be reviewed before drawing ROIs on the T 1 map. This issue is another reason why midwall, septum-only ROIs tend to be more reliable.
Free-breathing T 1 mapping acquisitions have been reported [34,[36][37][38][39], and aim to automatically exclude images with large misregistrations. While this may be more feasible for independent images, as used in SASHA [16], the impact of poor breath-holding or misgating variations on the later points of a Look-Locker set is convoluted. Furthermore, these methods can extend scan time, and often employ undersampling.

Cardiac triggering and cardiac motion
Source images for mapping must be acquired in the same phase of the cardiac cycle to ensure registration for pixelwise T 1 map calculation.
Some variation in R-R interval is normal, and MOLLIvariant sequences record the real-time R-R increments to the inversion-recovery time during each Look-Locker set. Arrhythmia may be tolerable provided the trigger to the pulse sequence is followed by a reasonably normal ventricular contraction and diastolic pause (diastasis). If arrhythmia interferes with diastolic timing, it is feasible to acquire the single-shot images in end-systole [66,67]. The end-systolic duration is less dependent on heart rate than diastasis, and may offer improved accuracy and precision in arrhythmia. Partial volume may also be less of an issue, as the contracted myocardium is thickened; however, the brevity of the end-systolic pause necessitates a shorter single-shot image readout, and thus spatial resolution is coarser.
A slightly shorter native T 1 has been reported for mapping at end systole versus end diastole [66,67]; however, this relationship flips after contrast administration, impacting ECV [68]. The cardiac phase of the image does not imply that the entire, usually slower, T 1 -relaxation process is sensitive to the myocardial relaxation or contraction state at the time of the image.
Similar to respiratory motion, cardiac mistriggering causes mis-registration of source images for T 1 mapping. Again, an elastic image registration algorithm may be able to account for this, but source images should be checked rigorously. If mis-registration goes uncorrected, distortion of fitted T 1 recovery curves is likely, particularly in the subendocardium. For this reason, some types of cardiac arrhythmia can be a major problem, but novel methods promise robust performance in such conditions [69].
The single-shot imaging duration should not exceed the length of the cardiac pause, be it diastolic or systolic. Tong et al. empirically estimated that the shot duration should not exceed 150 ms for minimal cardiac motion artefacts [70]. It is relatively straightforward to plan single-shot imaging to coincide with the required cardiac phase, as timings can be measured from a bSSFP cine acquired during routine setup, or may be semi-automated [71].
The image-readout duration can also be reduced using parallel imaging methods that acquire coil profiles separately: namely, before the scan or immediately after the last single-shot image. Note a related pitfall with coil profiles or any other prescan applied immediately prior to a T 1 mapping acquisition is that the longitudinal magnetisation may not have recovered before the first inversion [48], though this is usually avoided with a pause for breath hold instruction to the patient.

Flowing blood
Measurement of blood T 1 is important for the calculation of ECV. Complications in Look-Locker correction arise for blood that is at least partially replaced by fresh wash-in in the image slice for each cardiac cycle. Completely "fresh" blood, while still acted upon by the initial nonselective inversion, has not experienced previous shots of the current Look-Locker set since inversion and, therefore, does not require Look-Locker correction. However, even a normal heart ejects only around 55-75% of left ventricular blood per cycle [72], so the true situation is probably a complex mixture of different magnetisation histories in left-ventricular blood. Furthermore, in the extreme case, for later images of a native T 1 mapping acquisition the arriving blood may have experienced a distorted magnetisation preparation at some upstream location, despite application of optimised preparation pulses. This can occur in short bore scanner systems [5], or in unusual flow pathways following repairs of congenital heart defects.

Magnetisation transfer (MT)
When estimating T 1 in the myocardium and blood pool, it is also important to consider the MT phenomenon, which has been shown to influence the accuracy of T 1 mapping [5,47]. Exchange between free and bound water pools within the tissue of interest reduces the bSSFP signal [73]. For inversion-recovery-based T 1 mapping, the bound pool is mostly unaffected by the inversion pulse, so exchange during the long inversion-recovery delays distorts the shape of the T 1 recovery curve and introduces T 1 underestimation. The extent of this effect varies between tissues, and is substantially smaller in blood than in myocardium [5,74].
The MT effect can be allayed by using SASHA with a three-parameter curve-fit, at the expense of reduced T 1 estimation precision. It can also be mitigated in inversionrecovery T 1 mapping by use of lower-flip-angle excitation pulses and a longer imaging TR. Alternatively, the MT effect could be exploited in native T 1 mapping for greater disease discrimination in pathologies such as myocardial infarction (MI), ischaemia, and iron overload-all of which have all demonstrated MT.

Summary of technical pitfalls
Each technical and physiological challenge listed here may cause errors in T 1 and ECV mapping, increasing bias, scatter, or both. Cardiac and respiratory motion are particularly problematic, and strict quality control routines may help correct these where possible. At 3T, inhomogeneous B 0 and RF transmit fields also become prominent sources of error, which may be mitigated with appropriate B 0 shimming and RF transmit calibration. However, despite these prominent pitfalls, no issue dominates, and with considerable expertise, care, and attention to multiple factors, users can mitigate many sources of error, with the level of optimisation depending on their specific clinical or research applications.

Contrast agents
Gadolinium-based contrast agents (GBCAs) in myocardial T 1 mapping are subject to their own issues and controversies for deriving estimates of myocardial ECV. This section discusses the various assumptions made about GBCAs in T 1 and ECV mapping.

Assumptions relating to contrast agents
Several basic assumptions regarding GBCA estimation of ECV are stated here first, with further details later. Strictly, we assume that the contrast agent has identical relaxivities, r 1 , in myocardium and blood pool: where R 1,myo and R 1,blood are the changes in relaxation rates in myocardium and blood, respectively, and [Gd] represents the concentration of GBCA, typically in millimole/ litre units [75,76]. The change in relaxation rate, R 1 , is given as: If we assume that r 1,myo = r 1,blood , as stated above, then: where the ratio [Gd] myo [Gd] blood is the partition coefficient, λ [75,77].
(3) R 1,myo = r 1,myo [Gd] myo (4) R 1,blood = r 1,blood [Gd] blood , Secondly, we assume that the GBCA does not enter myocytes or blood cells, and that it is instead in dynamic equilibrium of water relaxation inside those cells, because of fast-exchange of water through cell walls, as follows: where Hct is the haematocrit. The term "fast" implies fast exchange of enough water across cell walls relative to the relevant T 1 range late after the myocardial first-pass.
Both amyloid deposition and collagen accumulation in fibrosis increase the interstitial space and break up myocyte packing. Collagen itself is of negligible non-permeated or "dark" volume (very short T 2 ), and is assumed to be highly permeable to interstitial fluid, including the GBCA. If the GBCA did not enter the collagen volume, but achieved relaxation equilibrium with it by fast exchange, it would manifest as an abnormally low ECV, resembling myocyte hypertrophy.
Finally, late after injection we assume that the concentration of GBCA in the interstitial fluid is equal to that in the blood plasma [77], and we calculate ECV as follows: as described by Messroghli et al. [78].

Contrast agent types
Several GBCAs are available for T 1 mapping applications, including gadopentetate dimeglumine (Gd-DTPA), gadobenate dimeglumine (Gd-BOPTA), and gadobutrol; these are known under the trade names "Magnevist", "Multihance", and "Gadovist", respectively. Each agent has differing relaxivities and binding properties, which can lead to differences in the estimated ECV; further complications arise due to relaxivity variations with different field strengths.
The Gd-BOPTA agent's aromatic ring enables weak plasma protein binding, leading to a lower molecular tumbling rate and thus a longer rotational MR correlation time and higher relaxivity in blood plasma and myocardial interstitial fluid compared to Gd-DTPA and gadobutrol. Furthermore, a lower dose of Gd-BOPTA has been shown to have similar diagnostic efficacy to a higher dose of Gd-DTPA in LGE imaging of MI [79]. Kawel et al. have shown that the use of Gd-DTPA leads to myocardial T 1 values around 15 ms lower than Gd-BOPTA, with no statistically significant difference seen in blood pool [80]. This results in slightly greater ECV values measured by Gd-DTPA, of the order of 0.01, perhaps due to Gd-BOPTA's binding to human serum albumin, which is responsible for its increased relaxivity. With regards to relaxivity variations with field strength, work by Rohrer et al. and Pintaske et al. has illustrated the variability in R 1 of GBCAs in human blood plasma for different contrast agent types and at different field strengths [81,82], which will lead to bias and variability in ECV measurements if not accounted for.
The relatively greater presence of albumin in blood versus myocardium means the distribution of protein-bound contrast agent between these pools is likely to be different than for non-protein-bound equivalents, altering the ratio of the change in relaxation rate of myocardium and blood and thus altering ECV. Given that this distorts one of the assumptions of in gadolinium-based ECV estimation, use of a protein-bound contrast agent will slightly modify partition coefficient estimation by T 1 mapping. If investigators plan to use one of these agents, they should do so consistently, and report this clearly in any inter-site comparisons.

Steady-state contrast versus bolus administration
The method of contrast administration also introduces variability to ECV calculation. Under most conditions, the twocompartment steady-state assumption, stated in Eq. (9), holds true for single bolus administration [83][84][85]. Early work in ECV used a primed infusion approach, whereby an initial loading bolus is followed by a continuous infusion of GBCA [86]. For most situations, the simpler bolus-only approach gives a similar ECV; however, for ECV greater than 0.4, in myocardial infarction (MI) and amyloidosis for example [87], it substantially overestimates ECV [84]. In a third method, several T 1 mapping acquisitions can be acquired during GBCA washout for improved accuracy in estimating the gadolinium blood-myocardium partition coefficient [24,88,89], which is estimated through the slope of a linear fit to myocardial R 1 versus blood R 1 . The partition coefficient has been shown to deviate from this model in the early washout phase [90,91], causing underestimation of ECV. This can be seen in Fig. 5, where data become markedly non-linear for blood R 1 values greater than 4 s −1 ; these data are often excluded from ECV calculations to avoid bias (Jerosch-Herold, personal communication).

Fast exchange assumption
Equations (7) and (8) are dependent on the fast exchange assumption, whereby we assume that water exchange between intracellular and extracellular compartments is sufficiently fast relative to the difference between the relaxation rates of the compartments considered in isolation [75]. In cases where this assumption is broken, where higher GBCA concentrations are used or post-GBCA measurements are made too early, as discussed above, ECV may be underestimated [92]. Regardless, in typical clinical T 1 -mapping situations, it appears that the fast-exchange assumption is sound.

Use of blood T 1 to calculate haematocrit
The blood for haematocrit assessment should be taken contemporaneously with T 1 mapping [13], to avoid unnecessary scatter in ECV. Synthetic haematocrit has recently been proposed as a means of streamlining ECV calculation [93]; it is calculated using the linear relationship between native blood T 1 and blood-analysed haematocrit. Support for this approach is spreading [93,94]; however, several issues have been identified that should be considered [95,96].

Summary of contrast agent issues
Although there are multiple issues with contrast agent types and field-strengths, none of these appear to dominate. There perhaps remains a dilemma between optimal ECV assessment and clinical feasibility: for example, the use of multiple acquisitions during GBCA washout, along with multiple averages for native T 1 scans, will improve ECV accuracy and precision, but such a protocol is difficult to fit into busy clinical schedules.

Errors in specific clinical applications
The aim of this section is to highlight the impact of the above topics in specific clinical applications, with examples and possible solutions.

Differences in age, sex, and myocardial region
There appear to be subtle differences in native myocardial T 1 related to sex and age, though there is currently no consensus on whether these also influence ECV [28,64,[97][98][99]. Several investigators have posited theories as to why native T 1 and ECV might increase or decrease with age, but the debate over these issues is outside the scope of this review. We should, however, point out that these changes are small, and demonstrable only over large groups, with a similar scatter to T 1 and ECV estimates in diffuse fibrosis, which are discussed later.
Regarding myocardial region, there appears to be no statistically significant difference between native T 1 measurements in the basal, mid, and apical regions of the left ventricle in healthy volunteers [100]. Several studies, however, have reported lower native T 1 values in the lateral wall versus the inter-ventricular septum [30,64,68,97,101]. It is likely this is mainly due to technical confounds rather than physiological differences, as cardiac motion, off-resonance by local B 0 distortion, and lower coil sensitivity all reduce accuracy and precision in the lateral wall. ECV, on the other hand, does not differ significantly between the lateral wall and septum [97], suggesting that reduction of native T 1 by off-resonance is likely the main source. Figure 6 illustrates the variation of native T 1 throughout the heart; lower T 1 values are seen in inferolateral segments, which typically demonstrate off-resonance due to interfaces with the lung and the posterior vein of the left ventricle [102,103].

Routine clinical applications of T 1 and ECV mapping
Myocardial T 1 and ECV mapping has attracted a lot of interest for both clinical and research applications due to its potential for accurate and precise tissue characterisation on a pixel-by-pixel basis. While T 1 mapping still shows promise for aiding precision medicine and influencing management of individual patients, recent work has been more pragmatic, with applications focusing on large patient populations and specific conditions.

Cardiac amyloidosis
Amyloid is a relatively rare multi-system condition where deposition of misfolded fibrillary protein in tissues and organs can cause expansion of the myocardial extracellular space and impairment of cardiac function [104]. This strongly increases native T 1 and ECV globally, with minimal overlap with healthy ranges [84], making them an excellent diagnostic tool [105]. Post-contrast T 1 mapping can also be helpful in highlighting abnormal GBCA washout kinetics, which show a specific pattern in amyloid patients [106], and a higher ECV in this context indicates a worse prognosis [107]. With large global changes and no concerns regarding myocardial region, this combination enables T 1 mapping to deliver strong diagnostic and prognostic utility.

Anderson-Fabry disease
Another rare multi-system syndrome, Anderson-Fabry disease (AFD) is characterised by intracellular accumulation of glycosphingolipids. This leads to progressive cardiac, renal, and cerebrovascular disease [108], and thus early diagnosis is extremely important for timely intervention. AFD markedly reduces global myocardial native T 1 compared to healthy volunteers, and thus represents a strong application of native T 1 mapping [109][110][111], again without concerns about regional myocardial differences. For several reasons, the reduced global native T 1 in AFD likely does not result directly from the short T 1 of fat [109], contrasting with apparent local T 1 increases sometimes seen in fatty infiltration of chronic MI, which is discussed later.

Myocarditis and Takotsubo cardiomyopathy
Myocarditis and Takotsubo cardiomyopathy are also potential clinical applications for T 1 mapping [112,113], being characterised by myocardial oedema, among several other markers. Oedema can be clearly highlighted on native T 1 maps due to increased interstitial fluid content; indeed, native T 1 mapping has higher diagnostic accuracy for identifying oedema than T 2 -weighted imaging in Takotsubo [113,114], and myocarditis [115]. In this application, although native T 1 is strongly increased, this is often sharply localised in myocardium, and thus operators should take care to localise to the relevant myocardial region. Furthermore, understanding 1 3 of localised technical pitfalls is also valuable: such as off-resonance, whose impact is further modulated by B 1 changes over the heart. It should be noted that published studies have typically excluded patients with major epicardial coronary disease or past MI. For T 1 mapping to become clinically meaningful in myocarditis and Takotsubo, it should not only positively confirm the diagnosis, but also rule out acute MI.

Potential clinical applications of T 1 mapping
In some pathologies, T 1 and ECV currently demonstrate limited sensitivity, but may ultimately be of clinical utility if their precision is improved, or if confounds are addressed.

Acute myocardial infarction
Like myocarditis and Takotsubo, acute MI also presents with oedema, which leads to elevated local T 1 on native T 1 maps [116][117][118]. There are, however, potential T 1 mapping pitfalls in acute MI, additional to that of local disease. Firstly, microvascular obstruction, or the "no reflow" phenomenon [119], causes derangement of the microvasculature, limiting blood flow post-reperfusion. This can distort an earlier assumption for ECV derivation, that of GBCA equilibrium between the infarct zone and the blood plasma. In particular, the necrotic core of the infarct will not be in equilibrium 15-20 min postbolus, requiring infusion for accurate ECV measurement [84].
Secondly, myocardial haemorrhage often occurs concomitantly with microvascular obstruction [120], and is characterised by extravasation of red blood cells through gaps in the endothelial walls. This leads to a cascade of haemoglobin decay products in the no reflow region during the weeks following reperfusion, with various iron states affecting T 2 , estimated T 1 , and true T 1 [121]. Clearly this poses problems for T 1 mapping, and thus CMR studies should be timed appropriately after reperfusion therapy [122].

Chronic myocardial infarction
For T 1 mapping, chronic MI presents the challenge of lipomatous metaplasia, which affects around 24-47% of MI patients [29,123]. This is characterised by fatty infiltration of myocardium, highlighting the following technical difficulty with fat partial volume in T 1 mapping.
Ignoring bSSFP characteristics, a mixture of in-phase lipid and water signals within a voxel produces a biexponential T 1 recovery curve. Attempting to fit these data with a monoexponential curve typically leads to lower myocardial T 1 estimates due to inclusion of short T 1 lipids, which have T 1 values around 370 ms at 1.5T and 450 ms at 3T [124]. However, in myocardial T 1 mapping the monoversus bi-exponential "model mismatch" issue is usually not the dominant factor, because most T 1 mapping uses bSSFP, in which fat and water are typically out of phase due to the frequency offset of fat [125]. Complex interference between fat and water signals leads to counterintuitive results: for smaller fat fractions, around 0.5-40% [29,126], an increase in the apparent myocardial T 1 occurs; higher fat fractions lead to undefined T 1 estimates; and fat-like T 1 values are seen only at the highest fat fractions. In chronic MI, fat fractions typically do not exceed 35% [29], and thus a positive T 1 estimation bias is expected, assuming fat and water are out of phase. This reduces the specificity of native T 1 mapping in chronic MI, because similar, but genuine, changes occur in oedema or inflammation. It is currently unclear what effect fatty changes have on ECV. See Fig. 7 for a plot of water and fat signals in bSSFP, along with a native T 1 map acquired in a chronic MI patient with lipomatous metaplasia.
The complex impact of fat partial volume in T 1 mapping might only reliably be reduced by replacing bSSFP with spoiled GRE imaging, in which the fat phase-difference depends purely on the TE. However, the reduced SNR of spoiled GRE relative to bSSFP would have to be considered. Alternatively, quantitative myocardial fat-fraction mapping could be applied [127], as recent work has incorporated fat-water separation into MOLLI and SASHA T 1 mapping in skeletal muscle [128] and the heart [32].

Diffuse myocardial fibrosis
Several pathologies cause diffuse fibrosis of the myocardium, such as hypertrophic and dilated cardiomyopathies (HCM and DCM), atrial fibrillation, aortic stenosis, heart failure with reduced or preserved ejection fractions (HFrEF and HFpEF), congenital heart disease, hypertension [129], and diabetes [130]. Certain drug therapies, such as alkylating agents in chemotherapy, can also lead to diffuse fibrosis [131], and possibly benefit from applications of T 1 mapping [132].
In general, native T 1 values are increased in diffuse fibrosis, but not to as great an extent as in oedema or amyloidosis, and ECV is also slightly elevated. In general, scatter in estimated T 1 values seems to be of a similar order to that seen in diffuse-fibrosis changes, and this has delayed the wider uptake of T 1 mapping. Several investigators have shown the usefulness of ECV as a prognostic indicator [133,134], but there seems to be little progress in taking this towards a per-patient test. In many applications there is a mixture of focal and diffuse fibrosis, which begs the question whether visible focal fibrosis was excluded from ROIs used in diffuse fibrosis studies [135], and even if so, what "mesoscopic" or "microscar" subvoxel focal fibrosis, or other myocardial changes, might be included in the socalled diffuse fibrosis assessment by T 1 [136] (Fig. 8). Fig. 7 The effect of intramyocardial lipids on native T 1 estimation in chronic myocardial infarction. The balanced steady-state free-precession off-resonance response is shown for myocardium and fat (a), for a modified Look-Locker inversion recovery protocol with a repetition time of 2.8 ms and a flip-angle of 35°. When fat and water are out of phase, lipids typically show an elevated T 1 in T 1 maps; when they are in-phase, lipid T 1 typically appears lower than myocardial T 1 . An example native T 1 map (b), acquired at 1.5T in a chronic myocardial infarction patient with lipomatous metaplasia, shows a reduced apparent T 1 associated with lipid signals (white arrow). Both the accuracy and precision of native T 1 estimates are influenced by this effect. Clearly there is substantial overlap between native T 1 and ECV values measured in controls versus those measured in patients with likely diffuse fibrosis. This is partly due to the small changes seen in diffuse fibrosis, particularly early in the disease when reversing it would be of great clinical benefit before irreversible damage occurs to the myocardium. It is also related to the many sources of dispersion in T 1 parameter estimation. As yet, it appears that there is no one factor that can be adjusted to achieve adequate T 1 estimation precision for detecting early diffuse fibrosis. Many recently reported clinical studies naturally retained older T 1 mapping protocols, due to the constraints of their study length, or follow-up periods in the "prognosis" papers. Despite these problems on the individual level, which may be overcome by stricter control of errors, T 1 mapping of diffuse fibrosis offers concrete benefits in large-scale studies: offering a means of testing treatment effects and characterising differences on a population level [130,137].

Summary of clinical sources of error
Both native T 1 and ECV are important clinical measures that allow us to characterise the myocardium in a fashion complementary to LGE. The substantial overlap of these measures between patients and controls in some cardiac conditions offers challenges to the clinical use of T 1 mapping, at least with current methodology. However, in specific conditions, such as amyloid and Anderson-Fabry disease, T 1 mapping has an important diagnostic and prognostic role, and is being widely adopted into routine clinical use.

Future directions of T 1 and ECV mapping
Potential future solutions for errors in T 1 and ECV mapping have been discussed throughout this review. We will now highlight several promising areas of development that may determine the future directions of the field.
Free-breathing T 1 mapping would appear to offer major gains with regards to T 1 mapping accuracy and precision, as well as enabling scanning of patients with compromised breath-holding. There are already several publications demonstrating the benefits of free-breathing T 1 mapping [36][37][38][39]; however, the added time required for these methods will limit their wider uptake, unless retrospective image registration can be robustly applied.
Certain recent implementations of myocardial T 1 mapping have incorporated simulations to improve the accuracy of bSSFP MOLLI [41,138,139], to enable spoiled GRE T 1 mapping [59], and to reduce the number of pause intervals between Look-Locker sets [140]. As yet, these methods do not offer an advantage in precision over the original MOLLI implementation, but further work may demonstrate benefits to their use [138].
Simultaneous T 1 , T 2 , and proton density mapping of the myocardium is also possible [141][142][143][144], with some methods being feasible in a single breath-hold [142][143][144]. Magnetic resonance fingerprinting is an extension of simulation-based methods that can offer T 1 , T 2 , and proton density maps [145], as well as other parameters modelled in the dictionary. It has recently been adapted to the heart [146]; however, further work is required, as currently it shows Fig. 8 Structural remodelling of myocardium in hypertension. Subvoxel heterogeneity can be seen in cardiomyocyte size, which ranges from hypertrophied to atrophied, and in fibrosis, which consists of microscopic scars, and perivascular and interstitial fibrosis. Reproduced with permission from Weber et al. [136] inferior precision to conventional T 1 mapping and requires long computation times for pattern matching.

Sacrificing T 1 accuracy for increased clinical utility
Placing a particular T 1 estimate in a local normal range may necessitate high precision, but not high accuracy. Indeed, accuracy may be sacrificed deliberately to yield T 1 estimates that are better able to discriminate normal tissue from pathology. For example, increasing the MOLLI excitation flip angle to 50° increases sensitivity to off-resonance, which reduces T 1 accuracy, but also increases MT effects, which differ between tissue types. This appears effective in detecting diffuse fibrosis in the septum [30,147]. The accuracy of this approach is limited, but the "true" T 1 is less relevant if local normal ranges are used, as they should be for any T 1 mapping implementation, under current guidelines [13].
Conversely, accuracy is important for establishing myocardial T 1 and ECV as clinical biomarkers using normal ranges that are transferable between sites and vendors [148].

Summary
This review has discussed reasons for inaccuracy and imprecision in T 1 and ECV mapping, and it should be clear from these considerations that users should take great care when deviating from manufacturers' advised T 1 mapping protocols. While improving the apparent quality of maps and source images, users may inadvertently reduce the precision of T 1 estimates, damaging clinical utility. Whatever T 1 mapping setup is used, it is essential that its performance is characterised in local normal ranges, and that it is applied only for those clinical questions that its precision can support. Ongoing quality control and reassessment is also required to ensure a local normal range remains valid; the reader is referred to the upcoming SCMR and EACVI consensus for recommendations in this regard.
In the research and clinical applications described here, current native T 1 and ECV mapping methods show utility in groupwise comparisons through to individual clinical tests. For some conditions, like diffuse fibrosis, mapping methods serve as weakly prognostic biomarkers that might be beneficial in combination with other diagnostic information about an individual patient. In other, albeit quite rare, conditions native T 1 and ECV mapping can provide strong diagnostic data.
Fundamentally, cardiac mapping methods have not changed radically since Messroghli et al. first introduced MOLLI [1]; however, their diversity offers challenges to inter-centre use, and new developments and quality controls are still evolving. New approaches that incorporate various forms of undersampling and modelling, such as fingerprinting, do not currently offer substantial gains in accuracy and precision, other than avoiding dependence on breath-holding in some cases.
It remains unclear how much of the scatter observed in T 1 estimates is due to physiological differences in true T 1 , or how much might be eliminated if the potentially correctable issues discussed here could be addressed robustly. If these pitfalls can be accounted for simply, quickly, and reliably, without need for specialist attention, T 1 and ECV mapping may ultimately support more widespread clinical applications.

Compliance with ethical standards
Conflict of interest The authors declare that they have no conflict of interest.
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.