Let’s talk about cardiac T1 mapping

Background Recent reports have shown that T1 mapping sequences agree in phantoms, but exhibit significant differences in vivo. To characterize these differences in the heart, one needs to consider the effects of magnetization transfer (MT) and the T2 relaxation time in the most commonly used cardiac T1 mapping sequences (MOLLI, ShMOLLI and SASHA). Methods Six explanted pig hearts were scanned weekly over a period of six weeks on a 3T system with the MOLLI, ShMOLLI, SASHA sequences and an inversion recovery sequence as reference. The T1 bias was computed as the difference between MOLLI, ShMOLLI, SASHA and the reference T1 values. We applied robust correlation statistics to assess the relationships between T1, T2 and MT. All data are publicly available at: http://neuropoly.pub/pigHeartsData. Results A systematic T1 bias was present for all sequences, with MOLLI and ShMOLLI underestimating T1 and SASHA slightly overestimating T1 compared to the reference. The correlation of T1 bias with T2 was weak and insignificant. However, MT showed significant associations with T1 bias for all sequences. Our analysis is also available at: http://neuropoly.pub/pigHeartsInteractive. Conclusion We investigated cardiac T1 mapping sequences in a setting that allowed us to explore their accuracy and their dependence on T2 and MT effects. The T2 effects were not significant, and could not explain the T1 bias of MOLLI, ShMOLLI, SASHA with respect to the reference. On the other hand, the T1 biases exhibited a strong correlation with MT. We conclude that inaccuracies in cardiac T1 mapping are primarily due to magnetization transfer.


BACKGROUND
Cardiac T1 mapping has the potential to play an important role in the diagnosis of heart disease 1 .
Abnormal native T1 values are indicators of infarction and inflammation 2 , protein deposition 3 , lipids 4 , and iron accumulation 5 . Abnormal post contrast T1 values are also consistent with extracellular volume expansion 6 . T1 mapping in the heart is primarily performed via inversion recovery 7 or saturation recovery sequences 8 , or a combination of the two 9 . The modified Look-Locker sequence (MOLLI) 10 and its shortened version ShMOLLI 11 use magnetization inversion, and then sample it multiple times during its return to equilibrium. The saturation recovery sequence SASHA 12 uses saturation pulses (90 degrees) to minimize the longitudinal magnetization, and then samples it once during its return to equilibrium after saturation. All methods provide a T1 map from the same cardiac phase within a single breath hold. T1 values obtained in normal volunteers show a significant difference between inversion recovery (MOLLI/ShMOLLI) and saturation recovery (SASHA) sequences 13,14 . Bloch simulations to characterize the difference between the two methods point to the T2 parameter as a major bias that affects the MOLLI method and its variant ShMOLLI, whereas the SASHA sequence was minimally affected by T2 1 5 . Recent published Bloch simulations 16,17 found that in addition to T2, the magnetization transfer (MT) effect also introduces bias in MOLLI and its variant ShMOLLI, but not in SASHA. On the flipside, SASHA tends to produce noisier maps, seemingly trading off precision for accuracy 14 .
The Society for Cardiovascular Magnetic Resonance (SCMR) and the CMR workshop group of the European Society of Cardiology have issued a consensus statement which mentioned Bloch simulations and phantoms as the first step in validating T1 mapping sequences, but recommended further research to study MT, T2 and T2* effects in biological tissues, as they may affect the T1 estimates differently compared to simulations and phantoms 18 .
To study MT and T2 effects in a more realistic environment, we imaged ex vivo pig hearts (in addition to phantoms) and applied a robust statistical framework to characterize and interpret the differences we observe between phantoms and ex vivo measurements.

Phantoms
The phantom used in this study was prepared by Captur and colleagues 19 and is composed of 3ൈ3 array tubes with different agar concentrations (ranging from 0.2 to 3 %) and different T1/T2 values. T1 values in the phantom range from 250 to 3000 ms and T2 values from 50 to 150 ms.
Only tubes with T2 values that correspond to the range of normal and abnormal myocardium T2 values were considered 20,21 , hence a total of 6 tubes with a range of T2s from 50 to 65 ms were considered for the analysis.

Ex vivo heart imaging
Six pig hearts were obtained from a local shop that supplies cardiology residents with specimens for anatomy lessons. The hearts were kept in ice during transportation, then immersed in saline for three hours before the first MRI scan. The hearts were imaged weekly (weeks 1, 2, 3) using a protocol approved by the institution's research ethics board and described below. Subsequently the hearts were immersed in 10% neutral buffered formalin solution (Chaptec Inc., QC, Canada) and imaged for another three weeks (weeks 4, 5, 6) using the same MRI protocol.

MRI protocol
T1, T2 and MTR measures were performed in a mid-ventricular slice of the ex-vivo hearts, and also in the agar phantoms using a 3T system (Magnetom Skyra, Siemens Healthcare, Erlangen, Germany) with an 18-channel phased-array cardiac coil. A simulated heart rate of 60 bpm was used. T1 measures were obtained in a single slice using the IR-TSE sequence as a reference, as well as the MOLLI, ShMOLLI and SASHA sequences (Myomaps, WIP1048 and WIP1041B respectively). Below is a more detailed description of the sequences used: • IR-TSE 22

Image analysis
For the ex vivo hearts, T1, T2 and MTR values were reported after manually delineating the endocardial and epicardial contours and generating masks of the LV myocardium on the parametric map using CVI42 (Circle CVI Inc., Calgary, Canada). All maps were resampled and aligned to match the MOLLI, ShMOLLI and SASHA outputs generated by the scanner.

Statistical Analysis
Scatter plots were inspected to evaluate whether bivariate distributions of T1 bias vs MTR and T1 bias vs T2 effects display a marked nonlinearity and discernable outliers. Relationships appeared to follow a monotonic linear trend, although salient outliers were present for some of the plots.
Moreover, some distributions exhibited heteroscedastic (fan-shaped) scatter. It is known that such distributions and outliers may pivot the linear regression line and misinterpret the strength of linear association between bivariate pairs. These effects can be mitigated using robust correlation analysis 24 .
Based on the initial inspection of data, an open source toolbox by Pernet  MTR, T1 bias). Finally, differences between paired correlations before and after fixation were tested for significance by calculating CIs using percentile bootstrapping.

Reproducibility
We believe that transparency is essential for the advancement of quantitative MRI, and that free dissemination of imaging data and analysis will help create a consensus in the field of qMRI, bringing us one step closer to clinical use. To that effect, we have made a strong effort to make all

Phantom
The correlations for the T1 bias against T2 and MTR in the agar phantom are shown in Figure 1,

Ex vivo
Upon 10% formalin fixation, a significant decrease was induced in the measured T1, T2 and MTR (Table 1). Although the SASHA T1bias exhibited a significant decrease of 73% after fixation, changes in MOLLI T1bias and ShMOLLI T1bias were substantially smaller (Fig. 2b). Fixation also caused a notable increase in the variability of myocardial T1 distributions for all sequences.     Fig. 4b and Table 2). Interestingly, while MTR was not correlated with the reference T1 before fixation (Table 1), formalin treatment gave rise to a significant correlation of 0.68 between IR T1 and MTR (Fig. 4d). The change between the paired correlations was also significant ( Table 2). Consistent with the phantom data, the ex vivo robust correlation analysis showed that the correlation between SASHA T1bias and T2 was weak and insignificant both before and after fixation ( Fig. 5a-b). Unlike the SASHA T1bias , the MOLLI T1bias and ShMOLLI T1bias exhibited much higher correlations with T2, also consistent with our phantom observations. While T2 did not account for the bulk of the T1 bias, MTR was highly correlated with T1 bias for all cardiac T1 mapping sequences (Fig. 6). More importantly, the formalin treatment only slightly reduced the correlations between T1 bias and MTR (see Fig. 6a-b) and none of the differences between these paired correlations were significant.

DISCUSSION
We investigated cardiac T1 mapping sequences in a setting that allowed us to explore their accuracy and their dependence on T2 and magnetization transfer effects. The T2 effects were not significant (Fig. 5), and could not account for the T1 bias of MOLLI, ShMOLLI, SASHA with respect to the reference. On the other hand, the T1 biases exhibited a strong correlation with MTR ( Fig. 6), and this correlation holds both before and after fixation. We conclude that inaccuracies in cardiac T1 mapping are primarily due to MT effects.
Differences between cardiac T1 mapping sequences have been reported before 14 , and these differences are often higher in vivo than in phantoms 14,25 . This is most likely due to the complex microstructure of human tissue, where MT effects are more complex. Formalin fixation gave us a sufficient dynamic range to explore realistic variations in the heart, but it also alters the biophysical properties of tissue 26,27 , which is why we cannot conclude that the same observations will hold in vivo and under pathological conditions.
We saw that the inversion recovery sequences (MOLLI and ShMOLLI) exhibited higher correlations with MT and T2, whereas SASHA, despite being a saturation recovery sequence, followed the reference measurement more closely. On the other hand, the raw SASHA images have lower SNR, resulting in less uniform T1 maps, suggesting a trade-off between accuracy and precision. Both need to be taken into account before deciding on a specific T1 mapping protocol.
The correlation of T1 bias with MTR was prominent irrespective of the fixation state for all cardiac T1 mapping techniques. Following the significant decrease in MTR upon fixation ( Table   1), correlation of IR T1 with MTR changed from a negligible to a significant level (Fig.4).
However, a similar change in the correlation between T2 and IR T1 is not observed. Therefore, what remains after subtracting reference IR T1 from MOLLI, ShMOLLI and SASHA can be taken as a metric to evaluate the tradeoff between accuracy and MT sensitivity.
Given that relaxation processes and MT are naturally intertwined and subjected to complex alternations depending on the environment of water protons 28,29 , phantoms and Bloch simulations could serve as the starting point for bringing the sequences together. Bloch simulations by Robson et al. reported at least a 10% reduction of the MOLLI T1 due to MT linked effects 30 . They also showed that MOLLI is more sensitive to T2 and MT in comparison to SASHA. Our phantom and ex-vivo findings confirm this previously theoretically shown effect. Moreover, a recent study has drawn further attention to the MT effects in myocardial relaxometry quantification by combining MT informed simulations with phantom measurements 31 . In that study, Xanthis et al.
tailored the MOLLI sequence to make it more T2, but less MT sensitive, which increased T1 estimation accuracy when compared to conventional MOLLI. This improvement provides support to our claim that MT contributes to T1 accuracy more than T2. While these effects are less pronounced in phantoms, it is critical to account for them in vivo in order to ensure proper comparison across sites, protocols and vendors.

CONCLUSION
Our ex-vivo analysis brings us closer to a unified theory of cardiac T1 mapping, but additional histological, statistical and clinical analyses are necessary to create a T1 mapping consensus in the CMR community. We believe that transparency is essential for the advancement of quantitative MRI, and that free dissemination of imaging data and analysis will help create a consensus in the field, bringing us one step closer to routine use of qMRI in the clinic. This manuscript is our contribution to these efforts.